4406 entries. 94 themes. Last updated December 26, 2016.

# Artificial Intelligence / Machine Learning / Deep Learning Timeline

## Bayes's Theorem for Calculating Inverse Probabilities 1763

On April 7, 1761 Thomas Bayes, an English clergyman and mathematician, died at the age of 59. Two years after his death, his paper, entitled "An Essay Towards Solving a Problem in the Doctrine of ChancesThomas Bayes was published in the Philosophical Transactions of the Royal Society 53 (1763) 370-418. Bayes's paper enunciated Bayes's Theorem for calculating "inverse probabilities”—the basis for methods of extracting patterns from data in decision analysisdata mining, statistical learning machinesBayesian networksBayesian inference.

"Whereas the ordinary rules of probability address such problems as 'what is the probability of drawing a yellow marble, if you draw three marbles from a sack containing 10 yellow marbles and 90 white marbles,' a Bayesian might ask the question, 'if I draw five marbles from a sack, and one is yellow and four are white, what is the probable distribution of the marbles in the sack?'  The advantage of inverse probability is that predictions can be continually refined as experience accumulates, so that if you draw five more marbles, and they are all white, that will change the probability prediction (and drawing a blue marble would drastically alter the situation), but Bayes’ theorem can easily accommodate any and all new information.  Bayes wrote his classic paper, 'An Essay towards solving a Problem in the Doctrine of Chances,' sometime in the late 1740s, but he never published it, for reasons unknown. After his death, his friend Richard Price found the paper among Bayes’ effects, and Price sent it for publication to John Canton of the Royal Society of London (apparently modifying the original paper considerably), and it appeared in the Philosophical Transactions in 1763. No one paid it the slightest attention. Ten years later, the Frenchman Pierre Simon Laplace independently discovered the rules of inverse probability, and although he later learned about Bayes’ paper and gave him priority, for the next century and a half Laplace got most of the credit (when credit was given at all--most statisticians did not consider Bayesian methods to be reputable, since they often involved making hunches and using gut feelings).  It wasn't until 1950 that the famous geneticist and mathematician R.A. Fisher first applied Bayes’ name to the methods of inverse probability, and since then, Bayes’ reputation has been gradually restored" (William B. Ashworth, Jr., email received on April 7, 2014.)

Hook & Norman, Origins of Cyberspace (2002) no. 1.

(This entry was last revised on April 7, 2014.)

## Von Kempelen "Invents" the Chess-Playing Turk & Edgar Allan Poe Compares it to Babbage's Difference Engine No. 1 1769 – 1836

In 1769 Hungarian author and inventor Wolfgang von Kempelen (Johann Wolfgang Ritter von Kempelen de Pázmánd; Hungarian: Kempelen Farkas) built his chess-playing Turk, an automaton that purported to play chess. Although the machine displayed an elaborate gear mechanism, its cabinet actually concealed a man controlling the moves of the machine.

Von Kempelen's Turk became a commercial sensation, deceiving a very large number of people. It became the most famous, or the most notorious, automaton in history. It also must have been kind of an open secret within the professional chess community because over the years numerous chess masters were hired so that The Turk could challenge all comers with its chess skills. With a skilled concealed operator the Turk won most of the games played during its demonstrations around Europe and the Americas for nearly 84 years, playing and defeating many challengers including Napoleon Bonaparte and Benjamin Franklin. Although many had suspected the hidden human operator, the hoax was first revealed by the English engineer Robert Willis in his illustrated pamphlet, An Attempt to Analyse the Automaton Chess Player of Mr. de Kempelen. With an Easy Method of Imitating the Movements of the Celebrated Figure. . .  (London, 1821). The operator or operators working within the mechanism during Kempelen's original tour remain a mystery; however after the engineer Johann Nepomuk Mälzel purchased the device in 1804, and exhibited it first in Europe and in 1826 in America, the chess masters who secretly operated it included Johann Allgaier, Hyacinthe Henri Boncourt, Aaron Alexandre, William Lewis, Jacques Mouret, and William Schlumberger. In 1818, for a short time while Boncourt was the operator of the Turk, he caught the flu and his chess performance was rather poor, and he could not control his coughing which could be heard by spectators, creating a certain embarrassment to Mälzel who owned the machine. For this reason Mälzel added some noisy gears to the Turk, which had no other purpose than to cover any noise that might come from the operator.

One of the most insightful commentators on The Turk was the American writer, poet, editor, literary critic, and magazinist Edgar Allan Poe. who in April 1836 published in the Southern Literary Messenger issued from Richmond, Virginia "Maelzel's Chess Player." In this article on automata Poe provided a very closely reasoned explanation of the concealed human operation of von Kempelen's Turk, which Poe had seen exhibited in Richmond by Maelzel a few weeks earlier.

Poe also briefly compared von Kempelen's Turk to Babbage's Difference Engine No. 1, which was limited to the computation of short mathematical tables, suggesting essentially that if the Turk was fully automated and had the ability to use the results of one logical operation to make a decision about the next one—what was later called "conditional branching" —it would be far superior to Babbage's machine. This feature Babbage later designed into his Analytical Engine

Here is Poe's comparison of the two machines:

"But if these machines were ingenious, what shall we think of the calculating machine of Mr. Babbage? What shall we think of an engine of wood and metal which can not only compute astronomical and navigation tables to any given extent, but render the exactitude of its operations mathematically certain through its power of correcting its possible errors? What shall we think of a machine which can not only accomplish all this, but actually print off its elaborate results, when obtained, without the slightest intervention of the intellect of man? It will, perhaps, be said, in reply, that a machine such as we have described is altogether above comparison with the Chess-Player of Maelzel. By no means — it is altogether beneath it — that is to say provided we assume (what should never for a moment be assumed) that the Chess-Player is a pure machine, and performs its operations without any immediate human agency. Arithmetical or algebraical calculations are, from their very nature, fixed and determinate. Certain data being given, certain results necessarily and inevitably follow. These results have dependence upon nothing, and are influenced by nothing but the data originally given. And the question to be solved proceeds, or should proceed, to its final determination, by a succession of unerring steps liable to no change, and subject to no modification. This being the case, we can without difficulty conceive the possibility of so arranging a piece of mechanism, that upon starting it in accordance with the data of the question to be solved, it should continue its movements regularly, progressively, and undeviatingly towards the required solution, since these movements, however complex, are never imagined to be otherwise than finite and determinate. But the case is widely different with the Chess-Player. With him there is no determinate progression. No one move in chess necessarily follows upon any one other. From no particular disposition of the men at one period of a game can we predicate their disposition at a different period. Let us place the first move in a game of chess, in juxta-position with the data of an algebraical question, and their great difference will be immediately perceived. From the latter — from the data — the second step of the question, dependent thereupon, inevitably follows. It is modelled by the data. It must be thus and not otherwise. But from the first move in the game of chess no especial second move follows of necessity. In the algebraical question, as it proceeds towards solution, the certainty of its operations remains altogether unimpaired. The second step having been a consequence of the data, the [column 2:] third step is equally a consequence of the second, the fourth of the third, the fifth of the fourth, and so on, and not possibly otherwise, to the end. But in proportion to the progress made in a game of chess, is the uncertainty of each ensuing move. A few moves having been made, no step is certain. Different spectators of the game would advise different moves. All is then dependent upon the variable judgment of the players. Now even granting (what should not be granted) that the movements of the Automaton Chess-Player were in themselves determinate, they would be necessarily interrupted and disarranged by the indeterminate will of his antagonist. There is then no analogy whatever between the operations of the Chess-Player, and those of the calculating machine of Mr. Babbage, and if we choose to call the former a pure machine we must be prepared to admit that it is, beyond all comparison, the most wonderful of the inventions of mankind. Its original projector, however, Baron Kempelen, had no scruple in declaring it to be a "very ordinary piece of mechanism — a bagatelle whose effects appeared so marvellous only from the boldness of the conception, and the fortunate choice of the methods adopted for promoting the illusion." But it is needless to dwell upon this point. It is quite certain that the operations of the Automaton are regulated by mind, and by nothing else. Indeed this matter is susceptible of a mathematical demonstration, a priori. The only question then is of the manner in which human agency is brought to bear. Before entering upon this subject it would be as well to give a brief history and description of the Chess-Player for the benefit of such of our readers as may never have had an opportunity of witnessing Mr. Maelzel's exhibition."

Even though the machine intelligence exhibited by the Turk was an illusion, von Kempelen's automaton was much later viewed as an analog to efforts in computer chess and artificial intelligence.

(This entry was last revised on 12-27-2014.)

## Samuel Butler Novel "Erewhon" Describes Artificial Consciousness 1872

In 1872 Erewhon: or, Over the Range, a satirical utopian novel by the English writer Samuel Butler, was published anonymously in London. A notable aspect of this satire on aspects of Victorian society, expanded from letters that Butler originally published in the New Zealand newspaper, The Press, was that Erewhonians believed that machines were potentially dangerous and that Erewhonian society had undergone a revolution that destroyed most mechanical inventions. In the section of Butler's satire called "The Book of the Machines" Butler appears to have imagined the possiblity of machine consciousness, or artificial consciousness, and that machines could replicate themselves

## Torres y Quevedo Invents the First Decision-Making Automaton 1912 – 1915

In 1912 Spanish civil engineer and mathematician, and Director of the Laboratory of Applied Mechanics at the Ateneo Científico, Literario y Artístico de MadridLeonardo Torres y Quevedo built the first decision-making automaton — a chess-playing machine that pit the machine’s rook and king against the king of a human opponent.  Torres's machine, which he called El Ajedrecista (The Chessplayer) used electromagnets under the board to "play" the endgame rook and king against the lone king.

"Well, not precisely play. But the machine could, in a totally unassisted and automated fashion, deliver mate with King and Rook against King. This was possible regardless of the initial position of the pieces on the board. For the sake of simplicity, the algorithm used to calculate the positions didn't always deliver mate in the minimum amount of moves possible, but it did mate the opponent flawlessly every time. The machine, dubbed El Ajedrecista (Spanish for “the chessplayer”), was built in 1912 and made its public debut during the Paris World Fair of 1914, creating great excitement at the time. It used a mechanical arm to make its moves and electrical sensors to detect its opponent's replies." (http://www.chessbase.com/newsprint.asp?newsid=1799, accessed 10-31-2012).

The implications of Torres's machines were not lost on all observers. On November 6, 1915 Scientific American magazine in their Supplement 2079 pp. 296-298 published an illustrated article entitled "Torres and his Remarkable Automatic Devices. He Would Substitute Machinery for the Human Mind."

## Von Neumann Invents the Theory of Games 1928

In 1928 Hungarian-American mathematician, physicist, economist and polymath John von Neumann then working at Humboldt-Universität zu Berlin, published "Zur Theorie der Gesellschaftsspiele" in Mathematische Annalen, 100, 295–300. This paper "On the Theory of Parlor Games" propounded the minimax theorem, inventing the theory of games.

## Alan Turing Publishes "On Computable Numbers," Describing What Came to be Called the "Turing Machine" November 30, 1936

In issues dated November 30 and December 23, 1936 of the Proceedings of the London Mathematical Society English mathematician Alan Turing published "On Computable Numbers", a mathematical description of what he called a universal machine— an astraction that could, in principle, solve any mathematical problem that could be presented to it in symbolic form. Turing modeled the universal machine processes after the functional processes of a human carrying out mathematical computation. In the following issue of the same journal Turing published a two page correction to his paper.

Undoubtedly the most famous theoretical paper in the history of computing, "On Computable Numbers" is a mathematical description an imaginary computing device designed to replicate the mathematical "states of mind" and symbol-manipulating abilities of a human computer. Turing conceived of the universal machine as a means of answering the last of the three questions about mathematics posed by David Hilbert in 1928: (1) is mathematics complete; (2) is mathematics consistent; and (3) is mathematics decidable.

Hilbert's final question, known as the Entscheidungsproblem, concerns whether there exists a defiinite method—or, in the suggestive words of Turing's teacher Max Newman, a "mechanical process"—that can be applied to any mathematical assertion, and which is guaranteed to produce a correct decision as to whether that assertion is true. The Czech logician Kurt Gödel had already shown that arithmetic (and by extension mathematics) was both inconsistent and incomplete. Turing showed, by means of his universal machine, that mathematics was also undecidable.

To demonstrate this, Turing came up with the concept of "computable numbers," which are numbers defined by some definite rule, and thus calculable on the universal machine. These computable numbers, "would include every number that could be arrived at through arithmetical operations, finding roots of equations, and using mathematical functions like sines and logarithms—every number that could possibly arise in computational mathematics" (Hodges, Alan Turing: The Enigma [1983] 100). Turing then showed that these computable numbers could give rise to uncomputable ones—ones that could not be calculated using a definite rule—and that therefore there could be no "mechanical process" for solving all mathematical questions, since an uncomputable number was an example of an unsolvable problem.

From 1936 to 1938 Mathematician Alan Turing spent more than a year at Princeton University studying mathematical logic with Alonzo Church, who was pursuing research in recursion theory. In August 1936 Church gave Turing's idea of a "universal machine" the name "Turing machine." Church coined the term in his relatively brief review of "On Computable Numbers." With regard to Turing's proof of the unsolvability of Hilbert's Entscheidungsproblem, Church acknowledged that "computability by a Turing machine. . . has the advantage of making the identification with effectiveness in the ordinary (not explicitly defined) sense evident immediately—i.e. without the necessity of proving elementary theorems." Church working independently of Turing, had arrived at his own answer to the Entscheidungsproblem a few months earlier. Norman, From Gutenberg to the Internet, Reading 7.2.

Independently of Alan Turing, mathematician and logician Emil Post of the City College of New York developed, and published in October 1936, a mathematical model of computation that was essentially equivalent to the Turing machine. Intending this as the first of a series of models of equivalent power but increasing complexity, he titled his paper Formulation 1. This model is sometime's called "Post's machine" or a Post-Turing machine.

In 1937 Turing and John von Neumann had their first discussions about computing and what would later be called “artificial intelligence” (AI). Always interested in practical applications of computing as well as theory, also while at Princeton, in 1937, believing that war with Germany was inevitable, Turing built in an experimental electromechanical cryptanalysis machine capable of binary multiplication in a university machine shop. After returning to England, on September 4, 1939, the day after Britain and France declared war on Germany, Turing reported to the Government Code and Cypher SchoolBletchley Park, in the town of Bletchley, England.

♦ In June 2013 it was my pleasure to purchase the famous copy of the offprint of "On Computable Numbers" along with the offprint of "On Computable Numbers . . . A Correction" that Turing presented to the English philosopher R. B. Braithwaite. One of very few copies in existence of the offprint, and possibly the only copy in private hands, the offprint sold for £205,000.  It was a price record for any offprint on a scientific or medical subject, for any publication in the history of computing, and probably the highest price paid for any scientific publication issued in the twentieth century.

Norman, From Gutenberg to the Internet, Reading 7.1. Hook & Norman, Origins of Cyberspace (2002) No. 394.

(This entry was last revised on 12-31-2014.)

## McCulloch & Pitts Publish the First Mathematical Model of a Neural Network 1943

In 1943 American neurophysiologist and cybernetician of the University of Illinois at Chicago Warren McCulloch and self-taught logician and cognitive psychologist Walter Pitts published “A Logical Calculus of the ideas Imminent in Nervous Activity,” describing the McCulloch - Pitts neuron, the first mathematical model of a neural network.

Building on ideas in Alan Turing’s “On Computable Numbers”, McCulloch and Pitts's paper provided a way to describe brain functions in abstract terms, and showed that simple elements connected in a neural network can have immense computational power. The paper received little attention until its ideas were applied by John von Neumann, Norbert Wiener, and others.

## Von Neumann & Morgenstern Issue "The Theory of Games and Economic Behavior" 1944

In 1944 mathematician, physicist, and economist John von Neumann, and economist Oskar Morgenstern published The Theory of Games and Economic Behavior in Princeton at the University Press.

Quantitative mathematical models for games such as poker or bridge at one time appeared impossible, since games like these involve free choices by the players at each move, and each move reacts to the moves of other players. However, in the 1920s John von Neumann single-handedly invented game theory, introducing the general mathematical concept of "strategy" in a paper on games of chance (Mathematische Annalen 100 [1928] 295-320). This contained the proof of his "minimax" theorem that says "a strategy exists that guarantees, for each player, a maximum payoff assuming that the adversary acts so as to minimize that payoff." The "minimax" principle, a key component of the game-playing computer programs developed in the 1950s and 1960s by Arthur Samuel, Allen Newell, Herbert Simon, and others was more fully articulated and explored in The Theory of Games and Economic Behavior, co-authored by von Neumann and Morgenstern.

Game theory, which draws upon mathematical logic, set theory and functional analysis, attempts to describe in mathematical terms the decision-making strategies used in games and other competitive situations. The Von Neumann-Morgenstern theory assumes (1) that people's preferences will remain fixed throughout; (2) that they will have wide knowledge of all available options; (3) that they will be able to calculate their own best interests intelligently; and (4) that they will always act to maximize these interests. Attempts to apply the theory in real-world situations have been problematical, and the theory has been criticized by many, including AI pioneer Herbert Simon, as failing to model the actual decision-making process, which typically takes place in circumstances of relative ignorance where only a limited number of options can be explored.

Von Neumann revolutionized mathematical economics. Had he not suffered an early death from cancer in 1957, most probably he would have received the first Nobel Prize in economics. (The first Nobel prize in economics was awarded in 1969; it cannot be awarded posthumously.) Several mathematical economists influenced by von Neumann's ideas later received the Nobel Prize in economics.

Hook & Norman, Origins of Cyberspace (2002) no. 953.

## One of the First Studies of Pattern Recognition 1947

In 1947 American logician Walter Pitts and psychiatrist neuroscientist Warren S. McCulloch published "How we Know Universals. The Perception of Auditory and Visual Forms," Bulletin of Mathematical Biophysics 9 (1947) 127-147. In this expansion of McCulloch and Pitts' "A Logical Calculus of the Ideas Immanent in Nervous Activity" (1943) Pitts and McCulloch implemented their notions by showing how the anatomy of the cerebral cortex might accommodate the identification of form independent of its angular size in the image, and other such operations in perception.

## Alan Turing's Contributions to Artificial Intelligence July 1948 – 1950

In July and August 1948 Alan Turing wrote a report for the National Physical Laboratory entitled Intelligent Machinery. In the report he stated that a thinking machine should be given the blank mind of an infant instead of an adult mind filled with opinions and ideas. The report contained an early discussion of neural networks. Turing estimated that it would take a battery of programmers fifty years to bring this learning machine from childhood to adult mental maturity. The report was not published until 1968.

In September 1948 Turing joined the computer project at Manchester University as Deputy Director and chief programmer.

In 1950 Turing published Computing Machinery and Intelligence, in which he described the “Turing Test" for determining whether a machine is “intelligent.”

"Turing predicted that machines would eventually be able to pass the test; in fact, he estimated that by the year 2000, machines with around 100 MB of storage would be able to fool 30% of human judges in a five-minute test, and that people would no longer consider the phrase "thinking machine" contradictory. (In practice, from 2009-2012, the Loebner Prize chatterbot contestants only managed to fool a judge once, and that was only due to the human contestant pretending to be a chatbot.) He further predicted that machine learning would be an important part of building powerful machines, a claim considered plausible by contemporary researchers in artificial intelligence.

"In a 2008 paper submitted to 19th Midwest Artificial Intelligence and Cognitive Science Conference, Dr. Shane T. Mueller predicted a modified Turing Test called a "Cognitive Decathlon" could be accomplished within 5 years.

"By extrapolating an exponential growth of technology over several decades, futurist Ray Kurzweil predicted that Turing test-capable computers would be manufactured in the near future. In 1990, he set the year around 2020. By 2005, he had revised his estimate to 2029.

## Using Neural Networks for Word Sense Disambiguation 1998

In 1998 cognitive scientist / entrepreneur Jeffrey Stibel, physicist, psychologist, neural scientist James A. Anderson, and others from the Department of Cognitive and Linguistic Sciences at Brown University created a word sense disambiguator using George A. Miller's WordNet lexical database.

Stibel and others applied this technology in Simpli, "an early search engine that offered disambiguation to search terms. A user could enter in a search term that was ambiguous (e.g., Java) and the search engine would return a list of alternatives (coffee, programming language, island in the South Seas)."

"The technology was rooted in brain science and built by academics to model the way in which the mind stored and utilized language."

"Simpli was sold in 2000 to NetZero. Another company that leveraged the Simpli WordNet technology was purchased by Google and they continue to use the technology for search and advertising under the brand Google AdSense.

"In 2001, there was a buyout of the company and it was merged with another company called Search123. Most of the original members joined the new company. The company was later sold in 2004 to ValueClick, which continues to use the technology and search engine to this day" (Wikipedia article on Simpli, accessed 05-10-2009).

## The First "Advanced" or "Freestyle" or "Centaur" Chess Event June 1998

The first Advanced Chess event, in which each human player used a computer chess program to help him explore the possible results of candidate moves, was held in June 1998 in León, Spain. The match was played between Garry Kasparov, using the German chess program Fritz 5, and Veselin Topalov, using ChessBase 7.0. The analytical engines used, such as FritzHIARCS and Junior, were integrated into these two programs, and could have been called at a click of the mouse. It was a 6-game match, and it was arranged in advance that the players would consult the built-in million games databases only for the 3rd and 4th game, and would only use analytical engines without consulting the databases for the remaining games. The time available to each player during the games was 60 minutes. The match ended in a 3-3 tie.

Since the first event Advanced Chess matches were often called Freestyle chess, in which players can play without computer assistance, or can simply follow the directions of a computer program, or can play as a "centaur", listening to the moves advocated by the AI but occasionally overriding them. In 2014 the best Freestyle chess player was Intagrand, a team of humans and several different chess programs.

## A Model of Cortical Processing as an Electronic Circuit of 16 "Neurons" that Could Select and Amplify Input Signals Much Like the Cortex of the Mammalian Brain 2000

In 2000 a research team from the Institute of Neuroinformatics ETHZ/UNI Zurich; Bell Laboratories, Murray Hill, NJ; and the Department of Brain and Cognitive Sciences & Department of Electrical Engineering and Computer Science at MIT created an electrical circuit of 16 "neurons" that could select and amplify input signals much like the cortex of the mammalian brain.

"Digital circuits such as the flip-flop use feedback to achieve multi-stability and nonlinearity tor estore signals to logical levels, for example 0 and 1. Analogue feedback circuits are generally designed to operate linearly, so that signals are over a range, and the response is unique. By contrast, the response of cortical circuits to sensory stimulation can be both multistable and graded. We propose that the neocortex combines digital selection of an active set of neurons with analogue response by dynamically varying the postive feedback inherent in its recurrent connections. Strong postive feedback causes differential instabilities that drive the selection of a set of active neurons under the constraints embedded in the synaptic weights. Once selected, the active neurons generate weaker, stable feedback that provides analogue amplication of the input. Here we present our model of cortical processing as an electronic circuit that emulates this hybrid operation, and so is able to perform computations that are similar to stimulus selection, gain modulation and spatiotemporal pattern generation in the neocortex" (Abstract)

R. Hahnloser, R. Sarpeshkar, M. Mahowald, R.J. Douglas and S. Seung: "Digital selection and analog amplification co-exist in an electronic circuit inspired by neocortex", Nature 405 (2000) 947-951.

## Conceiving and Building a Machine-Readable Database to Present Information "Filtered, Selected and Presented According to the Needs of the Individual User" 2000 – 2007

In 2000 American inventor, scientist, engineer, entrepreneur, and author William Daniel "Danny" Hillis wrote a paper entitled Aristotle (The Knowledge Web). In 2007, at the time of founding Metaweb Technologies to develop aspects of ideas expressed in his Aristotle paper, Hillis wrote:

"In retrospect the key idea in the "Aristotle" essay was this: if humans could contribute their knowledge to a database that could be read by computers, then the computers could present that knowledge to humans in the time, place and format that would be most useful to them.  The missing link to make the idea work was a universal database containing all human knowledge, represented in a form that could be accessed, filtered and interpreted by computers.

"One might reasonably ask: Why isn't that database the Wikipedia or even the World Wide Web? The answer is that these depositories of knowledge are designed to be read directly by humans, not interpreted by computers. They confound the presentation of information with the information itself. The crucial difference of the knowledge web is that the information is represented in the database, while the presentation is generated dynamically. Like Neal Stephenson's storybook, the information is filtered, selected and presented according to the specific needs of the viewer. ["In his book Diamond Age, the science fiction writer Neil Stephenson describes an automatic tutor called The Primer that grows up with a child. Stephenson's Primer does everything described above and more. It becomes a friend and playmate to the heroine of the novel, and guides not only her intellectual but also her emotional development" (from Hillis's Aristotle, 2000).

"John, Robert and I started a project,  then a company, to build that computer-readable database. How successful we will be is yet to be determined, but we are really trying to build it:  a universal database for representing any knowledge that anyone is willing to share. We call the company Metaweb, and the free database, Freebase.com. Of course it has none of the artificial intelligence described in the essay, but it is a database in which each topic is connected to other topics by links that describe their relationship. It is built so that computers can navigate and present it to humans. Still very primitive, a far cry from Neal Stephenson's magical storybook, it is a step, I hope, in the right direction" (http://edge.org/conversation/addendum-to-aristotle-the-knowledge-web, accessed 02-02-2014).

## The Film: "A. I. Artificial Intelligence" 2001

Steven Spielberg

The movie poster for A.I. Artificial Intelligence

Stanley Kubrick

In 2001 American director, screen writer and film producer Steven Spielberg directed, co-authored and produced, through DreamWorks and Amblin Entertainment, the science fiction film A.I. Artificial Intelligence, telling the story of David, an android robot child programmed with the ability to love and to dream. The film explored the hopes and fears involved with efforts to simulate human thought processes, and the social consequences of creating robots that may be better than people at specialized tasks.

The film was a 1970s project of Stanley Kubrick, who eventually turned it over to Spielberg. The project languished in development hell for nearly three decades before technology advanced sufficiently for a successful production. The film required enormously complex puppetry, computer graphics, and make-up prosthetics, which are well-described and explained in the supplementary material in the two-disc special edition of the film issued on DVD in 2002.

## A More Efficient Way to Teach Individual Layers of Neurons for Deep Learning 2006

In the mid-1980s, British-born computer scientist and psychologist Geoffrey Hinton and others helped revive resarch interest in neural networks with so-called “deep” models that made better use of many layers of software neurons. But the technique still required major human intervention: programmers had to label data before feeding it to the network, and complex speech or image recognition required more computer power than was available.

During the first decade of the 21st century Hinton and colleagues at the University of Toronto made some fundamental conceptual breakthroughs that have led to advances in unsupervised learning procedures for neural networks with rich sensory input.

"In 2006, Hinton developed a more efficient way to teach individual layers of neurons. The first layer learns primitive features, like an edge in an image or the tiniest unit of speech sound. It does this by finding combinations of digitized pixels or sound waves that occur more often than they should by chance. Once that layer accurately recognizes those features, they’re fed to the next layer, which trains itself to recognize more complex features, like a corner or a combination of speech sounds. The process is repeated in successive layers until the system can reliably recognize phonemes or objects" (Robert D. Hof, "Deep Learning," MIY Technology Review, April 23, 2013, accessed 11-10-2014).

Hinton, G. E.; Osindero, S.; Teh, Y., "A fast learning algorithm for deep belief nets", Neural Computation 18 #7 (2006) 1527–1554.

## IBM Begins Development of Watson, the First Cognitive Computer 2007

David Ferrucci

The Watson deep question answering computing system lab

The Watson Research Center

In 2007 David Ferrucci, leader of the Semantic Analysis and Integration Department at IBM’s Watson Research Center, Yorktown Heights, New York,  and his team began development of Watson, a special-purpose computer system designed to push the envelope on deep question and answering, deep analytics, and the computer's understanding of natural language. "Watson" became the firstg cognitive computer, combinding machine learning and artificial intelligence.

## Checkers is "Solved" April 29, 2007

Jonathan Shaeffer with a checkers board after "solving" the game of checkers

The University of Alberta seal

Jonathan Schaeffer and his team at the University of Alberta announced on April 29, 2007 that the game of checkers was "solved". Perfect play led to a draw.

"The crucial part of Schaeffer's computer proof involved playing out every possible endgame involving fewer than 10 pieces. The result is an endgame database of 39 trillion positions. By contrast, there are only 19 different opening moves in draughts. Schaeffer's proof shows that each of these leads to a draw in the endgame database, providing neither player makes a mistake.

"Schaeffer was able to get his result by searching only a subset of board positions rather than all of them, since some of them can be considered equivalent. He carried out a mere 1014 calculations to complete the proof in under two decades. 'This pushes the envelope as far as artificial intelligence is concerned,' he says.

"At its peak, Schaeffer had 200 desktop computers working on the problem full time, although in later years he reduced this to 50 or so. 'The problem is such that if I made a mistake 10 years ago, all the work from then on would be wrong,' says Schaeffer. 'So I've been fanatical about checking for errors.' " (http://www.newscientist.com/article/dn12296-checkers-solved-after-years-of-number-crunching.html, accessed 01-24-2010).

Based on this proof, Schaeffer's checkers-playing program Chinook, could no longer be beaten. The best an opponent could hope for is a draw.

## The SyNAPSE Neuromorphic Machine Technology Project Begins 2008

Traditional stored-program von Neumann computers are constrained by physical limits, and require humans to program how computers interact with their environments. In contrast the human brain processes information autonomously, and learns from its environment. Neuromorphic electronic machines— computers that function more like a brain— may enable autonomous computational solutions for real-world problems with many complex variables. In 2008 DARPA awarded the first funding to HRL Laboratories, Hewlett-Packard and IBM Research for SyNAPSE (Systems of Neuromorphic Adaptive Plastic Scalable Electronics)—an attempt to build a new kind of cognitive computer with form, function and architecture similar to the mammalian brain. The program sought to create electronic systems inspired by the human brain that could understand, adapt and respond to information in ways fundamentally different from traditional computers.

"The initial phase of the SyNAPSE program developed nanometer scale electronic synaptic components capable of adapting the connection strength between two neurons in a manner analogous to that seen in biological systems (Hebbian learning), and simulated the utility of these synaptic components in core microcircuits that support the overall system architecture" (Wikipedia article on SyNAPSE, accessed 10-20-2013).

## Using Automation to Find "Fundamental Laws of Nature" April 3, 2009

Michael Schmidt and Hod Lipson of Cornell University published "Distilling Free-Form Natural Laws from Experimental Data," Science 3 April 2009: Vol. 324. no. 5923, pp. 81 - 85 DOI: 10.1126/science.1165893.  The paper described a computer program that sifted raw and imperfect data to uncover fundamental laws of nature.

"For centuries, scientists have attempted to identify and document analytical laws that underlie physical phenomena in nature. Despite the prevalence of computing power, the process of finding natural laws and their corresponding equations has resisted automation. A key challenge to finding analytic relations automatically is defining algorithmically what makes a correlation in observed data important and insightful. We propose a principle for the identification of nontriviality. We demonstrated this approach by automatically searching motion-tracking data captured from various physical systems, ranging from simple harmonic oscillators to chaotic double-pendula. Without any prior knowledge about physics, kinematics, or geometry, the algorithm discovered Hamiltonians, Lagrangians, and other laws of geometric and momentum conservation. The discovery rate accelerated as laws found for simpler systems were used to bootstrap explanations for more complex systems, gradually uncovering the "alphabet" used to describe those systems" (Abstract from Science)

## Robot Scientist becomes the First Machine to Discover New Scientific Knowledge April 3, 2009

Ross D. King, Jem Rowland and 11 co-authors from the Department of Computer Science at Aberystwyth University, Aberystwyth, Wales, and the University of Cambridge, published "The Automation of Science," Science 3 April 2009: Vol. 324. no. 5923, pp. 85 - 89 DOI: 10.1126/science.1165620. In this paper they described a Robot Scientist which the researchers believed was the first machine to have independently discovered new scientific knowledge. The robot, called Adam, was a computer system that fully automated the scientific process.

"Prof Ross King, who led the research at Aberystwyth University, said: 'Ultimately we hope to have teams of human and robot scientists working together in laboratories'. The scientists at Aberystwyth University and the University of Cambridge designed Adam to carry out each stage of the scientific process automatically without the need for further human intervention. The robot has discovered simple but new scientific knowledge about the genomics of the baker's yeast Saccharomyces cerevisiae, an organism that scientists use to model more complex life systems. The researchers have used separate manual experiments to confirm that Adam's hypotheses were both novel and correct" (http://www.eurekalert.org/pub_releases/2009-04/babs-rsb032709.php).

"The basis of science is the hypothetico-deductive method and the recording of experiments in sufficient detail to enable reproducibility. We report the development of Robot Scientist "Adam," which advances the automation of both. Adam has autonomously generated functional genomics hypotheses about the yeast Saccharomyces cerevisiae and experimentally tested these hypotheses by using laboratory automation. We have confirmed Adam's conclusions through manual experiments. To describe Adam's research, we have developed an ontology and logical language. The resulting formalization involves over 10,000 different research units in a nested treelike structure, 10 levels deep, that relates the 6.6 million biomass measurements to their logical description. This formalization describes how a machine contributed to scientific knowledge" (Abstract in Science).

## The TV Show "Jeopardy" Provides a Good Model of the Semantic Analysis and Integration Problem April 22, 2009

On April 22, 2009 David Ferrucci, leader of the Semantic Analysis and Integration Department at IBM's T. J. Watson's Research Center, and Eric Nyberg, and several co-authors published the IBM Research Report: Towards the Open Advancement of Question Answering Systems.

Section 4.2.3. of the report included an analysis of why the television game show Jeopardy! provided a good model of the semantic analysis and integration problem.

## IBM's Watson Question Answering System Challenges Humans at "Jeopardy" April 27, 2009

On April 27, 2009 IBM announced that its Watson Question Answering (QA) System will challenge humans in the television quiz show Jeopardy!

"IBM is working to build a computing system that can understand and answer complex questions with enough precision and speed to compete against some of the best Jeopardy! contestants out there.

"This challenge is much more than a game. Jeopardy! demands knowledge of a broad range of topics including history, literature, politics, film, pop culture and science. What's more, Jeopardy! clues involve irony, riddles, analyzing subtle meaning and other complexities at which humans excel and computers traditionally do not. This, along with the speed at which contestants have to answer, makes Jeopardy! an enormous challenge for computing systems. Code-named "Watson" after IBM founder Thomas J. Watson, the IBM computing system is designed to rival the human mind's ability to understand the actual meaning behind words, distinguish between relevant and irrelevant content, and ultimately, demonstrate confidence to deliver precise final answers.

"Known as a Question Answering (QA) system among computer scientists, Watson has been under development for more than three years. According to Dr. David Ferrucci, leader of the project team, 'The confidence processing ability is key to winning at Jeopardy! and is critical to implementing useful business applications of Question Answering.

"Watson will also incorporate massively parallel analytical capabilities and, just like human competitors, Watson will not be connected to the Internet, or have any other outside assistance.

"If we can teach a computer to play Jeopardy!, what could it mean for science, finance, healthcare and business? By drastically advancing the field of automatic question answering, the Watson project's ultimate success will be measured not by daily doubles, but by what it means for society" (http://www.research.ibm.com/deepqa/index.shtml, accessed 06-16-2010).

On June 16, 2010 The New York Times Magazine published a long article by Clive Thompson on IBM's Watson's challenge of humans in Jeopardy! entitled, in the question response language of Jeopardy!, "What is I.B.M.'s Watson?."

♦ In December 2013 answers to frequently asked questions concerning Watson and Jeopardy! were available from IBM's website at this link.

## Wolfram/Alpha is Launched May 16, 2009

On May 16, 2009 Stephen Wolfram and Wolfram Research, Champaign, Illinois, launched Wolfram|Alpha, a computational data engine with a new approach to knowledge extraction, based on natural language processing, a large library of algorithms, and an NKS (New Kind of Science) approach to answering queries.

The Wolfram|Alpha engine differed from traditional search engines in that it did not simply return a list of results based on a query, but instead computed an answer.

## An Algorithm to Decipher Ancient Texts September 2, 2009

"Researchers in Israel say they have developed a computer program that can decipher previously unreadable ancient texts and possibly lead the way to a Google-like search engine for historical documents.

"The program uses a pattern recognition algorithm similar to those law enforcement agencies have adopted to identify and compare fingerprints.

"But in this case, the program identifies letters, words and even handwriting styles, saving historians and liturgists hours of sitting and studying each manuscript.

"By recognizing such patterns, the computer can recreate with high accuracy portions of texts that faded over time or even those written over by later scribes, said Itay Bar-Yosef, one of the researchers from Ben-Gurion University of the Negev.

" 'The more texts the program analyses, the smarter and more accurate it gets,' Bar-Yosef said.

"The computer works with digital copies of the texts, assigning number values to each pixel of writing depending on how dark it is. It separates the writing from the background and then identifies individual lines, letters and words.

"It also analyses the handwriting and writing style, so it can 'fill in the blanks' of smeared or faded characters that are otherwise indiscernible, Bar-Yosef said.

"The team has focused their work on ancient Hebrew texts, but they say it can be used with other languages, as well. The team published its work, which is being further developed, most recently in the academic journal Pattern Recognition due out in December but already available online. A program for all academics could be ready in two years, Bar-Yosef said. And as libraries across the world move to digitize their collections, they say the program can drive an engine to search instantaneously any digital database of handwritten documents. Uri Ehrlich, an expert in ancient prayer texts who works with Bar-Yosef's team of computer scientists, said that with the help of the program, years of research could be done within a matter of minutes. 'When enough texts have been digitized, it will manage to combine fragments of books that have been scattered all over the world,' Ehrlich said" (http://www.reuters.com/article/newsOne/idUSTRE58141O20090902, accessed 09-02-2009).

## Google Introduces Google Goggles December 8, 2009

On Cember 8, 2009 Google introduced Google Goggles image recognition and search technology for the Android mobile device operating system.  If you photographed certain types of individual objects with your mobile phone the program would recognize them and automatically display links to relevant information on the Internet.If you pointed your phone at a building the program would identify it by GPS and identify it. Then if you clicked on the name of the building it would bring up relevant Internet links.

♦ On May 7, 2010 you could watch a video describing the features of Google Goggles at this link:

http://www.google.com/mobile/goggles/#text

## "The Never-Ending Language Learning System" January 2010

Supported by DARPA and Google, in January 2010 Tom M. Mitchell and his team at Carnegie Mellon University initiated the Never-Ending Language Learning System, or NELL, in an effort to develop a method for machines to teach themselves semantics, or the meaning of language.

"Few challenges in computing loom larger than unraveling semantics, understanding the meaning of language. One reason is that the meaning of words and phrases hinges not only on their context, but also on background knowledge that humans learn over years, day after day" (http://www.nytimes.com/2010/10/05/science/05compute.html?_r=1&hpw).

"NELL has been in continuous operation since January 2010. For the first 6 months it was allowed to run without human supervision, learning to extract instances of a few hundred categories and relations, resulting in a knowledge base containing approximately a third of a million extracted instances of these categories and relations. At that point, it had improved substantially its ability to read three quarters of these categories and relations (with precision in the range 90% to 99%), but it had become inaccurate in extracting instances of the remaining fourth of the ontology (many had precisions in the range 25% to 60%).

"The estimated precision of the beliefs it had added to its knowledge base at that point was 71%. We are still trying to understand what causes it to become increasingly competent at reading some types of information, but less accurate over time for others. Beginning in June, 2010, we began periodic review sessions every few weeks in which we would spend about 5 minutes scanning each category and relation. During this 5 minutes, we determined whether NELL was learning to read it fairly correctly, and in case not, we labeled the most blatant errors in the knowledge base. NELL now uses this human feedback in its ongoing training process, along with its own self-labeled examples. In July, a spot test showed the average precision of the knowledge base was approximately 87% over all categories and relations. We continue to add new categories and relations to the ontology over time, as NELL continues learning to populate its growing knowledge base" (http://rtw.ml.cmu.edu/rtw/overview, accessed 10-06-2010).

## "The World's First Full-Size Robotic Girlfriend" January 9, 2010

On January 9, 2010 Artificial intelligence engineer Douglas Hines of TrueCompanion.com introduced Roxxxy at the AVN Adult Entertainment Expo in Las Vegas, Nevada.

" 'She doesn't vacuum or cook, but she does almost everything else,' said her inventor, Douglas Hines, who unveiled Roxxxy last month at the Adult Entertainment Expo in Las Vegas, Nevada.

"Lifelike dolls, artificial sex organs and sex-chat phone lines have been keeping the lonely company for decades. But Roxxxy takes virtual companionship to a new level. Powered by a computer under her soft silicone skin, she employs voice-recognition and speech-synthesis software to answer questions and carry on conversations. She even comes loaded with five distinct 'personalities,' from Frigid Farrah to Wild Wendy, that can be programmed to suit customers' preferences.

" 'There's a tremendous need for this kind of product,' said Hines, a computer scientist and former Bell Labs engineer. Roxxxy won't be available for delivery for several months, but Hines is taking pre-orders through his Web site, TrueCompanion.com, where thousands of men have signed up. 'They're like, 'I can't wait to meet her,' ' Hines said. 'It's almost like the anticipation of a first date.' Women have inquired about ordering a sex robot, too. Hines says a female sex therapist even contacted him about buying one for her patients.

"Roxxxy has been like catnip to talk-show hosts since her debut at AEE, the largest porn-industry convention in the country. In a recent monologue, Jay Leno expressed amazement that a sex robot could carry on lifelike conversations and express realistic emotions. 'Luckily, guys,' he joked, 'there's a button that turns that off.' Curious conventioneers packed Hines' AEE booth last month in Las Vegas, asking questions and stroking Roxxxy's skin as she sat on a couch in a black negligee.

" 'Roxxxy generated a lot of buzz at AEE,' said Grace Lee, spokeswoman for the porn-industry convention. 'The prevailing sentiment of everyone I talked to about Roxxxy is 'version 1.0,' but people were fascinated by the concept, and it caused them to rethink the possibilities of 'sex toys.' '

"Hines, a self-professed happily married man from Lincoln Park, New Jersey, says he spent more than three years developing the robot after trying to find a marketable application for his artificial-intelligence technology. Roxxxy's body is made from hypoallergenic silicone -- the kind of stuff in prosthetic limbs -- molded over a rigid skeleton. She cannot move on her own but can be contorted into almost any natural position. To create her shape, a female model spent a week posing for a series of molds. The robot runs on a self-contained battery that lasts about three hours on one charge, Hines says. Customers can recharge Roxxxy with an electrical cord that plugs into her back.

"A motor in her chest pumps heated air through a tube that winds through the robot's body, which Hines says keeps her warm to the touch. Roxxxy also has sensors in her hands and genital areas -- yes, she is anatomically correct -- that will trigger vocal responses from her when touched. She even shudders to simulate orgasm. When someone speaks to Roxxxy, her computer converts the words to text and then uses pattern-recognition software to match them against a database containing hundreds of appropriate responses. The robot then answers aloud -- her prerecorded 'voice' is supplied by an unnamed radio host -- through a loudspeaker hidden under her wig.

" 'Everything you say to her is processed. It's very near real time, almost without delay,' Hines said of the dynamics of human-Roxxxy conversation. 'To make it as realistic as possible, she has different dialogue at different times. She talks in her sleep. She even snores.' (The snoring feature can be turned off, he says.) Roxxxy understands and speaks only English for now, but Hines' True Companion company is developing Japanese and Spanish versions. For an extra fee, he'll also record customizable dialogue and phrases for each client, which means Roxxxy could talk to you about NASCAR, say, or the intricacies of politics in the Middle East" (http://www.cnn.com/2010/TECH/02/01/sex.robot/, accessed 02-06-2010).

In December 2013 I revisited the Truecompanion.com website, which then advertised Roxxxy as "World's First Sex Robot: Always Turned on and Ready to Talk or Play." By then the company had diversified into three models of female sex robots, and was planning to introduce Rocky, a male sex robot: "Rocky is described as everyone's dream date! – just imagine putting together a great body along with a sparkling personality where your man is focused on making you happy!"

## The First Fragment of Contemporary Classical Music Composed by a Computer in its Own Style October 15, 2010

On October 15, 2010 the Iamus computer cluster developed by Francisco Vico and associates at the Universidad de Málaga, using the Melomics system, composed Opus One. This composition was arguably the first fragment of professional contemporary classical music ever composed by a computer in its own style, rather than emulating the style of existing composers.

"Melomics (derived from the genomics of melodies) is a propietary computational system for the automatic composition of music (with no human intervention), based on bioinspired methods and commercialized by Melomics Media" (Wikipedia article on Melomics, accessed 11-13-2013).

## Kinect for Xbox is Introduced November 4, 2010

On November 4, 2010 Microsoft introduced Kinect, a natural user interface providing full-body 3D motion capture, facial recognition, and voice recognition, for the Xbox 360 video game platform. The device featured an "RGB camera, depth sensor and multi-array microphone running proprietary software."  It enabled users to control and interact with the Xbox 360 without the need to touch a game controller.

"The system tracks 48 parts of your body in three-dimensional space. It doesn’t just know where your hand is, like the Wii. No, the Kinect tracks the motion of your head, hands, torso, waist, knees, feet and so on" (http://www.nytimes.com/2010/11/04/technology/personaltech/04pogue.html?scp=1&sq=kinect&st=cse, accessed 11-04-2010).

## Can an Artificial Intelligence Get into the University of Tokyo? 2011

In 2011 National Institute of Informatics in Japan initiated the Todai Robot Project with the goal of achieving a high score on the National Center Test for University Admissions by 2016, and passing the University of Tokyo entrance exam in 2021.

"INTERVIEW WITH Yusuke Miyao, June 2013

Associate Professor, Digital Content and Media Sciences Research Division, NII; Associate Professor, Department of Informatics; "Todai Robot Project" Sub-Project Director

Can a Robot Get Into the University of Tokyo?
The Challenges Faced by the Todai Robot Project

Tainaka　Could you tell us the objectives of the project?

Miyao　We are researching the process of thinking by developing a computer program that will be able to pass the University of Tokyo entrance exam. The program will need to integrate multiple artificial intelligence technologies, such as language understanding, in order to develop all of the processes, from reading the question to determining the correct answer. While the process of thinking is first-nature to people, many of the processes involved in mental computation are still mysteries, so the project will be taking on challenges that previous artificial intelligence research has yet to touch.

Tainaka　You're not going to making a physical robot?

Miyao　No. What we'll be making is a robot brain. It won't be an actual robot that walks through the gate, goes to the testing site, picks up a pencil, and answers the questions.

Tainaka　Why was passing the university entrance exam selected as the project's goal?

Miyao　The key point is that what's difficult for people is different than what's difficult for computers. Computers excel at calculation, and can beat professional chess and shogi players at their games. IBM's "Watson" question-answering system*1 became a quiz show world champion. For a person, beating a professional shogi player is far harder than passing the University of Tokyo entrance exam, but for a computer, shogi is easier. What makes the University of Tokyo entrance exam harder is that the rules are less clearly defined than they are for shogi or a quiz show. From the perspective of using knowledge and data to answer questions, the university entrance exam requires a more human-like approach to information processing. However, it does not rely as much on common sense as an elementary school exam or everyday life, so it's a reasonable target for the next step in artificial intelligence research.
Tainaka　Elementary school exam questions are more difficult?

Miyao　For example, consider the sentence "Assuming there is a factory that can build 3 cars per day, how many days would it take to build 12 cars?" A computer would not be able to create a formula that expresses this in the same way a person could, near-instantaneously. It wouldn't understand the concepts of "car" or "factory", so it wouldn't be able to understand the relationship between them. Compared to that, calculating integrals is far easier.

Tainaka　The National Center Test for University Admissions is multiple choice, and the second-stage exam is a short answer exam, right?

Miyao　Of course, the center test is easier, and it has clear right and wrong answers, making it easier to grade. For the second-stage exam, examinees must give written answers, so during the latter half of the project, we will be shifting our focus on creating answers which are clear and comprehensible to human readers.

Tainaka　Does the difficulty vary by test subject?

Miyao　What varies more than the difficulty itself are the issues that have to be tackled by artificial intelligence research. The social studies questions, which test knowledge, rely on memory, so one might assume they would be easy for computers, but it's actually difficult for a computer to determine if the text of a problem corresponds to knowledge the computer possesses. What makes that identification possible is "Textual Entailment Recognition"*2, an area in which we are making progress, but still face many challenges. Ethics questions, on the other hand, frequently cover common sense, and require the reader to understand the Japanese language, so they are especially difficult for computers, which lack this common sense. Personally, I had a hard time with questions requiring memorization, so I picked ethics. (laughs)

Tainaka　So ethics and language questions are difficult because they involve common sense.

Miyao　Similar challenges are encountered with English, other than the common sense issue. For example, English questions include fill-in-the-blank questions, but it's difficult to pick natural conversational answers without actual life experience. Reading comprehension questions test logical and rational thought, but it's not really clear what this "logical and rational thought" consists of. The question, then, is how to teach "logical and rational thought" to computers. Also, for any subject, questions sometimes include photos, graphs, and comic strips. Humans understand them unconsciously, but it's extremely difficult to have computers understand them.

Tainaka　Aren't mathematical formula questions easy to answer?

Miyao　If they were presented as pure formulas, computers would excel at them, but the reality is not so simple. The questions themselves are written in natural language, making it difficult to map to the non-linguistic world of formulas. The same difficulty can be found with numerical fields, like physics or chemistry, or in fields which are difficult to convert into computer-interpretable symbols, such as the emotional and situational experience of reading a novel. That's what makes elementary school exams difficult.

Tainaka　There are a mountain of problems.

Miyao　There are many problems that nobody has yet taken on. That's what makes it challenging, and it's very exciting working with people from different fields. Looking at the practical results of this project, our discoveries and developments will be adapted for use in general purpose systems, such as meaning-based searching and conversation systems, real-world robot interfaces, and the like. The Todai Robot Project covers a diverse range of research fields, and NII plans to build an infrastructure, organizing data and creating platforms, and bring in researchers from both inside and outside Japan to achieve our objectives. In the future we will build an even more open platform, creating opportunities for members of the general public to participate as well, and I hope anyone motivated will take part" (http://21robot.org/introduce/NII-Interview/, accessed 12-30-2013).

## IBM's Watson Question Answering System Defeats Humans at Jeopardy! February 14 – February 16, 2011

LOn February 14, 2011 IBM's Watson question answering system supercomputer, developed at IBM's T J Watson Research Center, Yorktown Heights, New York, running DeepQA software, defeated the two best human Jeopardy! players, Ken Jennings and Brad Rutter. Watson's hardware consisted of 90 IBM Power 750 Express servers. Each server utilized a 3.5 GHz POWER7 eight-core processor, with four threads per core. The system operatesd with 16 terabytes of RAM.

The success of the machine underlines very significant advances in deep analytics and the ability of a machine to process unstructured data, and especially to intepret and speak natural language.

"Watson is an effort by I.B.M. researchers to advance a set of techniques used to process human language. It provides striking evidence that computing systems will no longer be limited to responding to simple commands. Machines will increasingly be able to pick apart jargon, nuance and even riddles. In attacking the problem of the ambiguity of human language, computer science is now closing in on what researchers refer to as the “Paris Hilton problem” — the ability, for example, to determine whether a query is being made by someone who is trying to reserve a hotel in France, or simply to pass time surfing the Internet.

"If, as many predict, Watson defeats its human opponents on Wednesday, much will be made of the philosophical consequences of the machine’s achievement. Moreover, the I.B.M. demonstration also foretells profound sociological and economic changes.

"Traditionally, economists have argued that while new forms of automation may displace jobs in the short run, over longer periods of time economic growth and job creation have continued to outpace any job-killing technologies. For example, over the past century and a half the shift from being a largely agrarian society to one in which less than 1 percent of the United States labor force is in agriculture is frequently cited as evidence of the economy’s ability to reinvent itself.

"That, however, was before machines began to 'understand' human language. Rapid progress in natural language processing is beginning to lead to a new wave of automation that promises to transform areas of the economy that have until now been untouched by technological change.

" 'As designers of tools and products and technologies we should think more about these issues,' said Pattie Maes, a computer scientist at the M.I.T. Media Lab. Not only do designers face ethical issues, she argues, but increasingly as skills that were once exclusively human are simulated by machines, their designers are faced with the challenge of rethinking what it means to be human.

"I.B.M.’s executives have said they intend to commercialize Watson to provide a new class of question-answering systems in business, education and medicine. The repercussions of such technology are unknown, but it is possible, for example, to envision systems that replace not only human experts, but hundreds of thousands of well-paying jobs throughout the economy and around the globe. Virtually any job that now involves answering questions and conducting commercial transactions by telephone will soon be at risk. It is only necessary to consider how quickly A.T.M.’s displaced human bank tellers to have an idea of what could happen" (John Markoff,"A Fight to Win the Future: Computers vs. Humans," http://www.nytimes.com/2011/02/15/science/15essay.html?hp, accessed 02-17-2011).

♦ As a result of this technological triumph, IBM took the unusal step of building a colorful website concerning all aspects of Watson, including numerous embedded videos.

♦ A few of many articles on the match published during or immediately after it included:

John Markoff, "Computer Wins on 'Jeopardy!': Trivial, It's Not," http://www.nytimes.com/2011/02/17/science/17jeopardy-watson.html?hpw

Samara Lynn, "Dissecting IBM Watson's Jeopardy! Game," PC Magazinehttp://www.pcmag.com/article2/0,2817,2380351,00.asp

John C. Dvorak, "Watson is Creaming the Humans. I Cry Foul," PC Magazinehttp://www.pcmag.com/article2/0,2817,2380451,00.asp

Henry Lieberman published a three-part article in MIT Technology Review, "A Worthwhile Contest for Artificial Intelligence" http://www.technologyreview.com/blog/guest/26391/?nlid=4132

♦ An article which discussed the weaknesses of Watson versus a human in Jeopardy! was Greg Lindsay, "How I Beat IBM's Watson at Jeopardy! (3 Times)" http://www.fastcompany.com/1726969/how-i-beat-ibms-watson-at-jeopardy-3-times

♦ An opinion column emphasizing the limitations of Watson compared to the human brain was Stanley Fish, "What Did Watson the Computer Do?" http://opinionator.blogs.nytimes.com/2011/02/21/what-did-watson-the-computer-do/

♦ A critical response to Stanley Fish's column by Sean Dorrance Kelly and Hubert Dreyfus, author of What Computers Can't Dowas published in The New York Times at: http://opinionator.blogs.nytimes.com/2011/02/28/watson-still-cant-think/?nl=opinion&emc=tya1

## The Impact of Automation on Legal Research March 4, 2011

"Armies of Expensive Lawyers Replaced by Cheaper Software," an article by John Markoff published in The New York Times, discussed the use of "e-discovery" (ediscovery) software which uses artificial intelligence to analyze millions of electronic documents from the linguistic, conceptual and sociological standpoint in a fraction of the time and at a fraction of the cost of the hundreds of lawyers previously required to do the task.

"These new forms of automation have renewed the debate over the economic consequences of technological progress.

"David H. Autor, an economics professor at the Massachusetts Institute of Technology, says the United States economy is being 'hollowed out.' New jobs, he says, are coming at the bottom of the economic pyramid, jobs in the middle are being lost to automation and outsourcing, and now job growth at the top is slowing because of automation.

" 'There is no reason to think that technology creates unemployment,' Professor Autor said. 'Over the long run we find things for people to do. The harder question is, does changing technology always lead to better jobs? The answer is no.'

"Automation of higher-level jobs is accelerating because of progress in computer science and linguistics. Only recently have researchers been able to test and refine algorithms on vast data samples, including a huge trove of e-mail from the Enron Corporation.

“ 'The economic impact will be huge,' said Tom Mitchell, chairman of the machine learning department at Carnegie Mellon University in Pittsburgh. 'We’re at the beginning of a 10-year period where we’re going to transition from computers that can’t understand language to a point where computers can understand quite a bit about language.'

## The Impact of Artificial Intelligence and Automation on Jobs March 6, 2011

In an op-ed column called Degrees and Dollars published in The New York Times Nobel Prize winning economist Paul Krugman of Princeton wrote concerning the impact of artificial intelligence and automation on jobs:

"The fact is that since 1990 or so the U.S. job market has been characterized not by a general rise in the demand for skill, but by “hollowing out”: both high-wage and low-wage employment have grown rapidly, but medium-wage jobs — the kinds of jobs we count on to support a strong middle class — have lagged behind. And the hole in the middle has been getting wider: many of the high-wage occupations that grew rapidly in the 1990s have seen much slower growth recently, even as growth in low-wage employment has accelerated."

"Some years ago, however, the economists David Autor, Frank Levy and Richard Murnane argued that this was the wrong way to think about it. Computers, they pointed out, excel at routine tasks, “cognitive and manual tasks that can be accomplished by following explicit rules.” Therefore, any routine task — a category that includes many white-collar, nonmanual jobs — is in the firing line. Conversely, jobs that can’t be carried out by following explicit rules — a category that includes many kinds of manual labor, from truck drivers to janitors — will tend to grow even in the face of technological progress.

"And here’s the thing: Most of the manual labor still being done in our economy seems to be of the kind that’s hard to automate. Notably, with production workers in manufacturing down to about 6 percent of U.S. employment, there aren’t many assembly-line jobs left to lose. Meanwhile, quite a lot of white-collar work currently carried out by well-educated, relatively well-paid workers may soon be computerized. Roombas are cute, but robot janitors are a long way off; computerized legal research and computer-aided medical diagnosis are already here.

"And then there’s globalization. Once, only manufacturing workers needed to worry about competition from overseas, but the combination of computers and telecommunications has made it possible to provide many services at long range. And research by my Princeton colleagues Alan Blinder and Alan Krueger suggests that high-wage jobs performed by highly educated workers are, if anything, more “offshorable” than jobs done by low-paid, less-educated workers. If they’re right, growing international trade in services will further hollow out the U.S. job market."

## The First Neurosynaptic Chips August 2011

In August 2011, as part of the SyNAPSE (Systems of Neuromorphic Adaptive Plastic Scalable Electronics) project, IBM researchers led by Dharmendra S. Modha, manager and lead researcher of the Cognitive Computing Group at IBM Almaden Research Center, demonstrated two neurosynaptic cores that moved beyond von Neumann architecture and programming to ultra-low, super-dense brain-inspired cognitive computing chips. These new silicon, neurosynaptic chips would be the building blocks for computing systems that emulate the brain's computing efficiency, size and power usage.

## Free Online Artificial Intelligence Course Attracts 58,000 Students August 15, 2011

Sebastian Thrun, Research Professor Computer Science at Stanford and a leading roboticist, and Peter Norvig, Director of Research at Google, Inc., in partnership with the Stanford University School of Engineering, offered a free online course entitled An Introduction to Artificial Intelligence

According to an article by John Markoff in The New York Times, by August 15, 2011 more than 58,000 students from around the world registered for this free course— nearly four times Stanford's entire student body.

"The online students will not get Stanford grades or credit, but they will be ranked in comparison to the work of other online students and will receive a 'statement of accomplishment.'

"For the artificial intelligence course, students may need some higher math, like linear algebra and probability theory, but there are no restrictions to online participation. So far, the age range is from high school to retirees, and the course has attracted interest from more than 175 countries" (http://www.nytimes.com/2011/08/16/science/16stanford.html?hpw, accessed 08-16-2011).

One fairly obvious reason why so many studients signed up is that Norvig is famous in the field as the co-author with Stuart Russell of the standard textbook on AI, Artificial Intelligence: A Modern Approach (first edition: 1995), which has been translated into many languages and has sold over 200,000 copies.

## Toward Cognitive Computing Systems August 18, 2011

On August 18, 2011 "IBM researchers unveiled a new generation of experimental computer chips designed to emulate the brain’s abilities for perception, action and cognition. The technology could yield many orders of magnitude less power consumption and space than used in today’s computers.

"In a sharp departure from traditional concepts in designing and building computers, IBM’s first neurosynaptic computing chips recreate the phenomena between spiking neurons and synapses in biological systems, such as the brain, through advanced algorithms and silicon circuitry. Its first two prototype chips have already been fabricated and are currently undergoing testing.

"Called cognitive computers, systems built with these chips won’t be programmed the same way traditional computers are today. Rather, cognitive computers are expected to learn through experiences, find correlations, create hypotheses, and remember – and learn from – the outcomes, mimicking the brains structural and synaptic plasticity.

"To do this, IBM is combining principles from nanoscience, neuroscience and supercomputing as part of a multi-year cognitive computing initiative. The company and its university collaborators also announced they have been awarded approximately $21 million in new funding from the Defense Advanced Research Projects Agency (DARPA) for Phase 2 of the Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) project. "The goal of SyNAPSE is to create a system that not only analyzes complex information from multiple sensory modalities at once, but also dynamically rewires itself as it interacts with its environment – all while rivaling the brain’s compact size and low power usage. The IBM team has already successfully completed Phases 0 and 1. " 'This is a major initiative to move beyond the von Neumann paradigm that has been ruling computer architecture for more than half a century,' said Dharmendra Modha, project leader for IBM Research. 'Future applications of computing will increasingly demand functionality that is not efficiently delivered by the traditional architecture. These chips are another significant step in the evolution of computers from calculators to learning systems, signaling the beginning of a new generation of computers and their applications in business, science and government.' " (http://www-03.ibm.com/press/us/en/pressrelease/35251.wss, accessed 08-21-2011). ## The First Complete Album Composed Solely by Computer and Recorded by Human Musicians September 2011 – July 2, 2012 In September 2011 the Iamus computer cluster developed by Francisco Vico and associates at the Universidad de Málaga produced a composition entitled Hello World! This classical clarinet-volin-piano trio was called the first full-scale work entirely composed by a computer without any human intervention, and automatically written in a fully-fledged score using conventional musical notation. Several months later, on July 2, 2012 four compositions by the Iamus computer premiered, and were broadcast live from the School of Computer Science at Universidad de Málaga, as one of the events included in the Alan Turing year. The compositions performed at this event were later recorded by the London Symphony Orchestra, and issued in 2012 as the album entitled Iamus. This compact disc was characterized by the New Scientist as the "first complete album to be composed solely by a computer and recorded by human musicians." Commenting on the authenticity of the music, Stephen Smoliar, critic of classical music at The San Francisco Examiner, wrote in a piece entitled "Thoughts about Iamus and the composition of music by computer," Examiner.com, January 4, 2013: "However, where listening is concerned, the method leading to the notation is secondary. What is primary is the act of making the music itself engaged by the performers and how the listener responds to what those performers do. Put another way, the music is in the performance, rather than in the composition without which that performance would not take place. The issue is not, as Smith seems to imply at the end of her BBC report, whether 'a computer could become a more prodigious composer than Mozart, Haydn, Brahms and Beethoven combined.' The computer is only prodigious at creating more documents, and what is most interesting about the documents generated by Iamus is their capacity to challenge the creative talents of performing musicians." ## The First Commercial Application of the IBM Watson Question Answering System: Medical Diagnostics September 12, 2011 Health Care insurance provider WellPoint, Inc. and IBM announced an agreement to create the first commercial applications of the IBM Watson question answering system. Under the agreement, WellPoint would develop and launch Watson-based solutions to help improve patient care through the delivery of up-to-date, evidence-based health care for millions of Americans, while IBM would develop the Watson healthcare technology on which WellPoint's solution will run. ## A Silicon Chip that Mimics How the Brain's Synapses Change in Response to New Information November 2011 In November 2011, a group of MIT researchers created the first computer chip that mimicked how the brain's neurons adapt in response to new information. This biological phenomenon, known as plasticity, is analog, ion-based communication in a synapse between two neurons. With about 400 transistors, the silicon chip can simulated the activity of a single brain synapse — a connection between two neurons that allows information to flow from one to the other. "There are about 100 billion neurons in the brain, each of which forms synapses with many other neurons. A synapse is the gap between two neurons (known as the presynaptic and postsynaptic neurons). The presynaptic neuron releases neurotransmitters, such as glutamate and GABA, which bind to receptors on the postsynaptic cell membrane, activating ion channels. Opening and closing those channels changes the cell’s electrical potential. If the potential changes dramatically enough, the cell fires an electrical impulse called an action potential. "All of this synaptic activity depends on the ion channels, which control the flow of charged atoms such as sodium, potassium and calcium. Those channels are also key to two processes known as long-term potentiation (LTP) and long-term depression (LTD), which strengthen and weaken synapses, respectively. " "The MIT researchers designed their computer chip so that the transistors could mimic the activity of different ion channels. While most chips operate in a binary, on/off mode, current flows through the transistors on the new brain chip in analog, not digital, fashion. A gradient of electrical potential drives current to flow through the transistors just as ions flow through ion channels in a cell. “ 'We can tweak the parameters of the circuit to match specific ion channels,” Poon says. 'We now have a way to capture each and every ionic process that’s going on in a neuron.' "Previously, researchers had built circuits that could simulate the firing of an action potential, but not all of the circumstances that produce the potentials. “If you really want to mimic brain function realistically, you have to do more than just spiking. You have to capture the intracellular processes that are ion channel-based,” Poon says" (http://www.mit.edu/newsoffice/2011/brain-chip-1115.html, accessed 01-01-2014). Rachmuth, G., Shouvai, H., Bear, M., Poon, C. "A biophysically-based neuromorphic model of spike rate- and timing-dependent plasticity," Proceedings of the National Academy of Sciences 108, no. 459, December 6, 2011, E1266-E1274, doi: 10.1073/pnas.1106161108 ## IBM's Watson Question Answering System to Team with Cedars-Sinai Oschin Comprehensive Cancer Institute December 16, 2011 Health Insurance provider WellPoint announced that the Cedars-Sinai Samuel Oschin Comprehensive Cancer Institute in Los Angeles would provide clinical expertise to help shape WellPoint's new health care solutions utilizing IBM's Watson question answering system. "It is estimated that new clinical research and medical information doubles every five years, and nowhere is this knowledge advancing more quickly than in the complex area of cancer care. "WellPoint believes oncology is one of the medical fields that could greatly benefit from this technology, given IBM Watson's ability to respond to inquiries posed in natural language and to learn from the responses it generates. The WellPoint health care solutions will draw from vast libraries of information including medical evidence-based scientific and health care data, and clinical insights from institutions like Cedars-Sinai. The goal is to assist physicians in evaluating evidence-based treatment options that can be delivered to the physician in a matter of seconds for assessment. WellPoint and Cedars-Sinai envision that this valuable enhancement to the decision-making process could empower physician-patient discussions about the best and most effective courses of treatment and improve the overall quality of patient care. "Cedars-Sinai was selected as WellPoint's partner based on its reputation as one of the nation's premier cancer institutions and its proven results in the diagnosis and treatment of complex cancers. Cedars-Sinai has experience and demonstrated success in working with technology innovators and shares WellPoint's commitment to improving the quality, efficiency and effectiveness of health care through innovation and technology. "Cedars-Sinai's oncology experts will help develop recommendations on appropriate clinical content for the WellPoint health care solutions. They will also assist in the evaluation and testing of the specific tools that WellPoint plans to develop for the oncology field utilizing IBM's Watson technology. The Cedars-Sinai cancer experts will enter hypothetical patient scenarios, evaluate the proposed treatment options generated by IBM Watson, and provide guidance on how to improve the content and utility of the treatment options provided to the physicians. "Leading Cedars-Sinai's efforts is M. William Audeh, M.D., medical director of its Samuel Oschin Comprehensive Cancer Institute. Dr. Audeh will work closely with WellPoint's clinical experts to provide advice on how the solutions may be best utilized in clinical practice to support increased understanding of the evolving body of knowledge in cancer, including emerging therapies not widely known by community physicians. As the solutions are developed, Dr. Audeh will also provide guidance on how the make the WellPoint offering useful and practical for physicians and patients. " 'As we design the WellPoint systems that leverage IBM Watson's capabilities, it is essential that we incorporate the highly-specialized knowledge and real-life practice experiences of the nation's premier clinical experts,' said Harlan Levine, MD, executive vice president of WellPoint's Comprehensive Health Solutions. 'The contributions from Dr. Audeh, coupled with the expertise throughout Cedars-Sinai's Samuel Oschin Comprehensive Cancer Institute, will be invaluable to implementing this WellPoint offering and could ultimately benefit millions of Americans across the country.' "WellPoint anticipates deploying their first offering next year, working with select physician groups in clinical pilots" (http://ir.wellpoint.com/phoenix.zhtml?c=130104&p=irol-newsArticle&ID=1640553&highlight=, accessed 12-17-2011). ### 2012 – 2016 ## A Large Scale Neural Network Appears to Emulate Activity in the Visual Cortex June 26, 2012 At the International Conference on Machine Learning held in Edinburgh, Scotland from June 26–July 1, 2012 researchers at Google and Stanford University reported that they developed software modeled on the way biological neurons interact with each other that taught itself to distinguish objects in ­YouTube videos. Although it was most effective recognizing cats and human faces, the system obtained 15.8% accuracy in recognizing 22,000 object categories from ImageNet, or 3,200 items in all, a 70 percent improvement over the previous best-performing software. To do so the scientists connected 16,000 computer processors to create a neural network for machine learning with more than one billion connections. Then they turned the neural network loose on the Internet to learn on its own. Having been presented with the experimental results before the meeting, on June 25, 2012 John Markoff published an article entitled "How Many Computers to Identify a Cat? 16,000," from which I quote selections" "Presented with 10 million digital images selected from YouTube videos, what did Google’s brain do? What millions of humans do with YouTube: looked for cats.... "The research is representative of a new generation of computer science that is exploiting the falling cost of computing and the availability of huge clusters of computers in giant data centers. It is leading to significant advances in areas as diverse as machine vision and perception, speech recognition and language translation. "Although some of the computer science ideas that the researchers are using are not new, the sheer scale of the software simulations is leading to learning systems that were not previously possible. And Google researchers are not alone in exploiting the techniques, which are referred to as “deep learning” models. Last year Microsoft scientists presented research showing that the techniques could be applied equally well to build computer systems to understand human speech.... "The [YouTube] videos were selected randomly and that in itself is an interesting comment on what interests humans in the Internet age. However, the research is also striking. That is because the software-based neural network created by the researchers appeared to closely mirror theories developed by biologists that suggest individual neurons are trained inside the brain to detect significant objects. "Currently much commercial machine vision technology is done by having humans 'supervise' the learning process by labeling specific features. In the Google research, the machine was given no help in identifying features. “ 'The idea is that instead of having teams of researchers trying to find out how to find edges, you instead throw a ton of data at the algorithm and you let the data speak and have the software automatically learn from the data,' Dr. Ng said. “ 'We never told it during the training, ‘This is a cat,’ ' said Dr. Dean, who originally helped Google design the software that lets it easily break programs into many tasks that can be computed simultaneously. 'It basically invented the concept of a cat. We probably have other ones that are side views of cats.' "The Google brain assembled a dreamlike digital image of a cat by employing a hierarchy of memory locations to successively cull out general features after being exposed to millions of images. The scientists said, however, that it appeared they had developed a cybernetic cousin to what takes place in the brain’s visual cortex. "Neuroscientists have discussed the possibility of what they call the 'grandmother neuron,' specialized cells in the brain that fire when they are exposed repeatedly or “trained” to recognize a particular face of an individual. “ 'You learn to identify a friend through repetition,' said Gary Bradski, a neuroscientist at Industrial Perception, in Palo Alto, Calif. "While the scientists were struck by the parallel emergence of the cat images, as well as human faces and body parts in specific memory regions of their computer model, Dr. Ng said he was cautious about drawing parallels between his software system and biological life. “ 'A loose and frankly awful analogy is that our numerical parameters correspond to synapses,' said Dr. Ng. He noted that one difference was that despite the immense computing capacity that the scientists used, it was still dwarfed by the number of connections found in the brain. “ 'It is worth noting that our network is still tiny compared to the human visual cortex, which is a million times larger in terms of the number of neurons and synapses,' the researchers wrote. "Despite being dwarfed by the immense scale of biological brains, the Google research provides new evidence that existing machine learning algorithms improve greatly as the machines are given access to large pools of data. “ 'The Stanford/Google paper pushes the envelope on the size and scale of neural networks by an order of magnitude over previous efforts,' said David A. Bader, executive director of high-performance computing at the Georgia Tech College of Computing. He said that rapid increases in computer technology would close the gap within a relatively short period of time: “The scale of modeling the full human visual cortex may be within reach before the end of the decade.” "Google scientists said that the research project had now moved out of the Google X laboratory and was being pursued in the division that houses the company’s search business and related services. Potential applications include improvements to image search, speech recognition and machine language translation. "Despite their success, the Google researchers remained cautious about whether they had hit upon the holy grail of machines that can teach themselves. “ 'It’d be fantastic if it turns out that all we need to do is take current algorithms and run them bigger, but my gut feeling is that we still don’t quite have the right algorithm yet,' said Dr. Ng. Quoc V. Le, Marc'Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg S. Corrado, Jeff DeanAndrew Y. Ng, "Building High-level Features Using Large Scale Upervised Learning," arXiv:1112.6209 [cs.LG] 12 July 2012. ## Using 100 Linked Computers and Artificial Intelligence to Re-Assemble Fragments from the Cairo Genizah May 2013 For years I have followed computer applications in the humanities. Some, such as From Cave Paintings to the Internet, are on a small personal scale. Others involve enormous corpora of data, as in computational linguistics, where larger seems always to be better. The project called "Re-joining the Cairo Genizah", a joint venture of Genazim, The Friedberg Genizah Project, founded in 1999 in Toronto, Canada, and The Blavatnik School of Computer Science at Tel-Aviv University, seems potentially to be one of the most productive large scale projects currently underway. Because about 320,000 pages and parts of pages from the Genizah — in Hebrew, Aramaic, and Judeo-Arabic (Arabic transliterated into Hebrew letters) — are scattered in 67 libraries and private collections around the world, only a fraction of them have been collated and cataloged. Though approximately 200 books were published on the Genizah manuscripts by 2013, perhaps only 4,000 of the manuscripts were pieced together through a painstaking, expensive, exclusive process that relied a lot on luck. In 2013 the Genazim project was underway to collate and piece together as many of these fragments as could be re-assembled using current computing technology: "First there was a computerized inventory of 301,000 fragments, some as small as an inch. Next came 450,000 high-quality photographs, on blue backgrounds to highlight visual cues, and a Web site where researchers can browse, compare, and consult thousands of bibliographic citations of published material. "The latest experiment involves more than 100 linked computers located in a basement room at Tel Aviv University here, cooled by standup fans. They are analyzing 500 visual cues for each of 157,514 fragments, to check a total of 12,405,251,341 possible pairings. The process began May 16 and should be done around June 25, according to an estimate on the project’s Web site. "Yaacov Choueka, a retired professor of computer science who runs the Friedberg-financed Genazim project in Jerusalem, said the goals are not only to democratize access to the documents and speed up the elusive challenge of joining fragments, but to harness the computer’s ability to pose new research questions. . . . "Another developing technology is a 'jigsaw puzzle' feature, with touch-screen technology that lets users enlarge, turn and skew fragments to see if they fit together. Professor Choueka, who was born in Cairo in 1936, imagines that someday soon such screens will be available alongside every genizah collection. And why not a genizah-jigsaw app for smartphones? “ 'The thing it really makes possible is people from all walks of life, in academia and out, to look at unpublished material,' said Ben Outhwaite, head of the Genizah Research Unit at Cambridge University, home to 60 percent of the fragments. 'No longer are we going to see a few great scholarly names hoarding particular parts of the genizah and have to wait 20 years for their definitive publication. Now everyone can dive in.' "What they will find goes far beyond Judaica. . . . Marina Rustow, a historian at Johns Hopkins University, said about 15,000 genizah fragments deal with everyday, nonreligious matters, most of them dated 950 to 1250. From these, she said, scholars learned that Cairenes imported sheep cheese from Sicily — it was deemed kosher — and filled containers at the bazaar with warm food in an early version of takeout" (http://www.nytimes.com/2013/05/27/world/middleeast/computers-piecing-together-jigsaw-of-jewish-lore.html?pagewanted=2&hp, accessed 05-27-2013) ## The Growing Economic and Social Impact of Artificial Intelligence December 29, 2013 On December 29, 2013 The New York Times published an article by Michael Fitzpatrick on Japan's Todai Robot Project entitled "Computers Jump to the Head of the Class." This was the first article that I ever read that spelled out the potential dystopian impact of advances in artificial intelligence on traditional employment and also on education. Because the article was relatively brief I decided to quote it in full: "TOKYO — If a computer could ace the entrance exam for a top university, what would that mean for mere mortals with average intellects? This is a question that has bothered Noriko Arai, a mathematics professor, ever since the notion entered her head three years ago. “I wanted to get a clear image of how many of our intellectual activities will be replaced by machines. That is why I started the project: Can a Computer Enter Tokyo University? — the Todai Robot Project,” she said in a recent interview. Tokyo University, known as Todai, is Japan’s best. Its exacting entry test requires years of cramming to pass and can defeat even the most erudite. Most current computers, trained in data crunching, fail to understand its natural language tasks altogether. Ms. Arai has set researchers at Japan’s National Institute of Informatics, where she works, the task of developing a machine that can jump the lofty Todai bar by 2021. If they succeed, she said, such a machine should be capable, with appropriate programming, of doing many — perhaps most — jobs now done by university graduates. With the development of artificial intelligence, computers are starting to crack human skills like information summarization and language processing. Given the exponential growth of computing power and advances in artificial intelligence, or A.I., programs, the Todai robot’s task, though daunting, is feasible, Ms. Arai says. So far her protégé, a desktop computer named Todai-kun, is excelling in math and history but needs more effort in reading comprehension. There is a significant danger, Ms. Arai says, that the widespread adoption of artificial intelligence, if not well managed, could lead to a radical restructuring of economic activity and the job market, outpacing the ability of social and education systems to adjust. Intelligent machines could be used to replace expensive human resources, potentially undermining the economic value of much vocational education, Ms. Arai said. “Educational investment will not be attractive to those without unique skills,” she said. Graduates, she noted, need to earn a return on their investment in training: “But instead they will lose jobs, replaced by information simulation. They will stay uneducated.” In such a scenario, high-salary jobs would remain for those equipped with problem-solving skills, she predicted. But many common tasks now done by college graduates might vanish. “We do not know in which areas human beings outperform machines. That means we cannot prepare for the changes,” she said. “Even during the industrial revolution change was a lot slower.” Over the next 10 to 20 years, “10 percent to 20 percent pushed out of work by A.I. will be a catastrophe,” she says. “I can’t begin to think what 50 percent would mean — way beyond a catastrophe and such numbers can’t be ruled out if A.I. performs well in the future.” She is not alone in such an assessment. A recent study published by the Program on the Impacts of Future Technology, at Oxford University’s Oxford Martin School, predicted that nearly half of all jobs in the United States could be replaced by computers over the next two decades. Some researchers disagree. Kazumasa Oguro, professor of economics at Hosei University in Tokyo, argues that smart machines should increase employment. “Most economists believe in the principle of comparative advantage,” he said. “Smart machines would help create 20 percent new white-collar jobs because they expand the economy. That’s comparative advantage.” Others are less sanguine. Noriyuki Yanagawa, professor of economics at Tokyo University, says that Japan, with its large service sector, is particularly vulnerable. “A.I. will change the labor demand drastically and quickly,” he said. “For many workers, adjusting to the drastic change will be extremely difficult.” Smart machines will give companies “the opportunity to automate many tasks, redesign jobs, and do things never before possible even with the best human work forces,” according to a report this year by the business consulting firm McKinsey. Advances in speech recognition, translation and pattern recognition threaten employment in the service sectors — call centers, marketing and sales — precisely the sectors that provide most jobs in developed economies. As if to confirm this shift from manpower to silicon power, corporate investment in the United States in equipment and software has never been higher, according to Andrew McAfee, the co-author of “Race Against the Machine” — a cautionary tale for the digitized economy. Yet according to the technology market research firm Gartner, top business executives worldwide have not grasped the speed of digital change or its potential impact on the workplace. Gartner’s 2013 chief executive survey, published in April, found that 60 percent of executives surveyed dismissed as “‘futurist fantasy” the possibility that smart machines could displace many white-collar employees within 15 years. “Most business and thought leaders underestimate the potential of smart machines to take over millions of middle-class jobs in the coming decades,” Kenneth Brant, research director at Gartner, told a conference in October: “Job destruction will happen at a faster pace, with machine-driven job elimination overwhelming the market’s ability to create valuable new ones.” Optimists say this could lead to the ultimate elimination of work — an “Athens without the slaves” — and a possible boom for less vocational-style education. Mr. Brant’s hope is that such disruption might lead to a system where individuals are paid a citizen stipend and be free for education and self-realization. “This optimistic scenario I call Homo Ludens, or ‘Man, the Player,’ because maybe we will not be the smartest thing on the planet after all,” he said. “Maybe our destiny is to create the smartest thing on the planet and use it to follow a course of self-actualization.” ## A Neural Network that Reads Millions of Street Numbers January 1, 2014 To read millions of street numbers on buildings photographed for Google StreetView, Google built a neural network that developed reading accuracy comparable to humans assigned to the task. The company uses the images to read house numbers and match them to their geolocation, storing the geolocation of each building in its database. Having the street numbers matched to physical location on a map is always useful, but it is particularly useful in places where street numbers are otherwise unavailable, or in places such as Japan and South Korea, where streets are rarely numbered in chronological order, but in other ways, such as the order in which they were constructed— a system that makes many buildings impossibly hard to find, even for locals. "Recognizing arbitrary multi-character text in unconstrained natural photographs is a hard problem. In this paper, we address an equally hard sub-problem in this domain viz. recognizing arbitrary multi-digit numbers from Street View imagery. Traditional approaches to solve this problem typically separate out the localization, segmentation, and recognition steps. In this paper we propose a unified approach that integrates these three steps via the use of a deep convolutional neural network that operates directly on the image pixels. We employ the DistBelief implementation of deep neural networks in order to train large, distributed neural networks on high quality images. We find that the performance of this approach increases with the depth of the convolutional network, with the best performance occurring in the deepest architecture we trained, with eleven hidden layers. We evaluate this approach on the publicly available SVHN dataset and achieve over 96% accuracy in recognizing complete street numbers. We show that on a per-digit recognition task, we improve upon the state-of-the-art and achieve 97.84% accuracy. We also evaluate this approach on an even more challenging dataset generated from Street View imagery containing several tens of millions of street number annotations and achieve over 90% accuracy. Our evaluations further indicate that at specific operating thresholds, the performance of the proposed system is comparable to that of human operators. To date, our system has helped us extract close to 100 million physical street numbers from Street View imagery worldwide." Ian J. Goodfellow, Yaroslav Bulatov, Julian Ibarz, Sacha Arnoud, Vinay Shet, "Multi-digit Number Recognition from Street ViewImagery using Deep Convolutional Neural Networks," arXiv:1312.6082v2. ## DeepFace, Facial Verification Software Developed at Facebook, Approaches Human Ability March 17, 2014 On March 17, 2014 MIT Technology Review published an article by Tim Simonite on Facebook's facial recognition software, DeepFace, which I quote: "Asked whether two unfamiliar photos of faces show the same person, a human being will get it right 97.53 percent of the time. New software developed by researchers at Facebook can score 97.25 percent on the same challenge, regardless of variations in lighting or whether the person in the picture is directly facing the camera. "That’s a significant advance over previous face-matching software, and it demonstrates the power of a new approach to artificial intelligence known as deep learning, which Facebook and its competitors have bet heavily on in the past year (see 'Deep Learning'). This area of AI involves software that uses networks of simulated neurons to learn to recognize patterns in large amounts of data. "'You normally don’t see that sort of improvement,' says Yaniv Taigman, a member of Facebook’s AI team, a research group created last year to explore how deep learning might help the company (see 'Facebook Launches Advanced AI Effort'). 'We closely approach human performance,' says Taigman of the new software. He notes that the error rate has been reduced by more than a quarter relative to earlier software that can take on the same task. "Facebook’s new software, known as DeepFace, performs what researchers call facial verification (it recognizes that two images show the same face), not facial recognition (putting a name to a face). But some of the underlying techniques could be applied to that problem, says Taigman, and might therefore improve Facebook’s accuracy at suggesting whom users should tag in a newly uploaded photo. "However, DeepFace remains purely a research project for now. Facebook released a research paper on the project last week, and the researchers will present the work at the IEEE Conference on Computer Vision and Pattern Recognition in June. 'We are publishing our results to get feedback from the research community,' says Taigman, who developed DeepFace along with Facebook colleagues Ming Yang and Marc’Aurelio Ranzato and Tel Aviv University professor Lior Wolf. "DeepFace processes images of faces in two steps. First it corrects the angle of a face so that the person in the picture faces forward, using a 3-D model of an 'average' forward-looking face. Then the deep learning comes in as a simulated neural network works out a numerical description of the reoriented face. If DeepFace comes up with similar enough descriptions from two different images, it decides they must show the same face. "The performance of the final software was tested against a standard data set that researchers use to benchmark face-processing software, which has also been used to measure how humans fare at matching faces. "Neeraj Kumar, a researcher at the University of Washington who has worked on face verification and recognition, says that Facebook’s results show how finding enough data to feed into a large neural network can allow for significant improvements in machine-learning software. 'I’d bet that a lot of the gain here comes from what deep learning generally provides: being able to leverage huge amounts of outside data in a much higher-capacity learning model,' he says. "The deep-learning part of DeepFace consists of nine layers of simple simulated neurons, with more than 120 million connections between them. To train that network, Facebook’s researchers tapped a tiny slice of data from their company’s hoard of user images—four million photos of faces belonging to almost 4,000 people. 'Since they have access to lots of data of this form, they can successfully train a high-capacity model,' says Kumar. ## IBM Launches "Watson Discovery Advisor" to Hasten Breakthroughs in Scientific and Medical Research August 27, 2014 On August 27, 2014 IBM launched Watson Discovery Advisor, a computer system that could quickly identify patterns in massive amounts of data, with the expectation that this system would hasten breakthroughs in science and medical research. The computer system, which IBM made available through the cloud, understood chemical compound interaction and human language, and could visually map out connections in data. The system used a number of computational techniques to deliver its results, including natural language processing, machine learning and hypothesis generation, in which a hypothesis is created and evaluated by a number of different analysis techniques. Baylor College of Medicine used the service to analyze 23 million abstracts of medical papers in order to find more information on the p53 tumor protein, in search of more information on how to turn it on or off. From these results, Baylor researchers identified six potential proteins to target for new research. Using traditional methods it typically took researchers about a year to find a single potentially useful target protein, IBM said. According to an article by Reuters published in The New York Times, "Some researchers and scientists have already been using Watson Discovery Advisor to sift through the sludge of scientific papers published daily. "Johnson & Johnson is teaching the system to read and understand trial outcomes published in journals to speed up studies of effectiveness of drugs. "Sanofi, a French pharmaceutical company is working with Watson to identify alternate uses for existing drugs. " 'On average, a scientist might read between one and five research papers on a good day,' said Dr. Olivier Lichtarge, investigator and professor of molecular and human genetics, biochemistry and molecular biology at Baylor College of Medicine. "He used Watson to automatically analyze 70,000 articles on a particular protein, a process which could have taken him nearly 38 years. “ 'Watson has demonstrated the potential to accelerate the rate and the quality of breakthrough discoveries,' he said." ## Three Breakthroughs that Finally Unleased AI on the World October 27, 2014 In "The Three Breakthroughs That Have Finally Unleased AI on the World", Wired Magazine, October 27, 2014, writer Kevin Kelly of Pacifica, California explained how breakthroughs in cheap parallel computation, big data, and better algorithms were enabling new AI-based services that were previously the domain of sci-fi and academic white papers. Within the near future AI would play greater and greater roles in aspects of everyday life, in products like Watson developed by IBM, and products from Google, Facebook and other companies. More significant than these observations were Kelly's views about the impact that these developments would have on our lives and how we may understand the difference between machine and human intelligence: "If AI can help humans become better chess players, it stands to reason that it can help us become better pilots, better doctors, better judges, better teachers. Most of the commercial work completed by AI will be done by special-purpose, narrowly focused software brains that can, for example, translate any language into any other language, but do little else. Drive a car, but not converse. Or recall every pixel of every video on YouTube but not anticipate your work routines. In the next 10 years, 99 percent of the artificial intelligence that you will interact with, directly or indirectly, will be nerdily autistic, supersmart specialists. "In fact, this won't really be intelligence, at least not as we've come to think of it. Indeed, intelligence may be a liability—especially if by “intelligence” we mean our peculiar self-awareness, all our frantic loops of introspection and messy currents of self-consciousness. We want our self-driving car to be inhumanly focused on the road, not obsessing over an argument it had with the garage. The synthetic Dr. Watson at our hospital should be maniacal in its work, never wondering whether it should have majored in English instead. As AIs develop, we might have to engineer ways to prevent consciousness in them—and our most premium AI services will likely be advertised as consciousness-free. "What we want instead of intelligence is artificial smartness. Unlike general intelligence, smartness is focused, measurable, specific. It also can think in ways completely different from human cognition. A cute example of this nonhuman thinking is a cool stunt that was performed at the South by Southwest festival in Austin, Texas, in March of this year. IBM researchers overlaid Watson with a culinary database comprising online recipes, USDA nutritional facts, and flavor research on what makes compounds taste pleasant. From this pile of data, Watson dreamed up novel dishes based on flavor profiles and patterns from existing dishes, and willing human chefs cooked them. One crowd favorite generated from Watson's mind was a tasty version of fish and chips using ceviche and fried plantains. For lunch at the IBM labs in Yorktown Heights I slurped down that one and another tasty Watson invention: Swiss/Thai asparagus quiche. Not bad! It's unlikely that either one would ever have occurred to humans. "Nonhuman intelligence is not a bug, it's a feature. The chief virtue of AIs will be their alien intelligence. An AI will think about food differently than any chef, allowing us to think about food differently. Or to think about manufacturing materials differently. Or clothes. Or financial derivatives. Or any branch of science and art. The alienness of artificial intelligence will become more valuable to us than its speed or power. . . . ## Google Develops A Neural Image Caption Generator to Translate Images into Words November 17, 2014 Having previously transformed the machine translation process by developing algorithms from vector space mathematics, in November 2014 Oriol Vinyals and colleagues at Google in Mountain View developed a neural image caption generator to translate images into words. Google's machine translation approach is: "essentially to count how often words appear next to, or close to, other words and then define them in an abstract vector space in relation to each other. This allows every word to be represented by a vector in this space and sentences to be represented by combinations of vectors. "Google goes on to make an important assumption. This is that specific words have the same relationship to each other regardless of the language. For example, the vector “king - man + woman = queen” should hold true in all languages. . . . "Now Oriol Vinyals and pals at Google are using a similar approach to translate images into words. Their technique is to use a neural network to study a dataset of 100,000 images and their captions and so learn how to classify the content of images. "But instead of producing a set of words that describe the image, their algorithm produces a vector that represents the relationship between the words. This vector can then be plugged into Google’s existing translation algorithm to produce a caption in English, or indeed in any other language. In effect, Google’s machine learning approach has learnt to “translate” images into words. To test the efficacy of this approach, they used human evaluators recruited from Amazon’s Mechanical Turk to rate captions generated automatically in this way along with those generated by other automated approaches and by humans. "The results show that the new system, which Google calls Neural Image Caption, fares well. Using a well known dataset of images called PASCAL, Neural image Capture clearly outperformed other automated approaches. “NIC yielded a BLEU score of 59, to be compared to the current state-of-the-art of 25, while human performance reaches 69,” says Vinyals and co" (http://www.technologyreview.com/view/532886/how-google-translates-pictures-into-words-using-vector-space-mathematics/, accessed 01-14-2015). Vinyals et al, "Show and Tell: A Neural Image Captional Generator" (2014) http://arxiv.org/pdf/1411.4555v1.pdf "Automatically describing the content of an image is a fundamental problem in artificial intelligence that connects computer vision and natural language processing. In thispaper we present a generative model based on a deep recurrent architecture that combines recent advances in computer vision and machine translation and that can be used to generate natural sentences describing an image. The model is trained to maximize the likelihood of the target description sentence given the training image. Experiments on several datasets show the accuracy of the model and the fluency of the language it learns solely from image descriptions. Our model is often quite accurate, which we verify both qualitatively and quantitatively. For instance, while the current state-of-the-art BLEU score (the higher the better) on the Pascal dataset is 25, our approach yields 59, to be compared to human performance around 69. We also show BLEU score improvements on Flickr30k, from 55 to 66, and on SBU, from 19 to 27" (Abstract). ## Skype Previews Skype Translator Real Time Translation Program December 15, 2014 On December 15, 2014 Skype, a division of Microsoft, announced the first phase of the Skype Translator preview program. In the first phase the program was available only for English and Spanish. The result of 15 years of work at Microsoft Research in deep learning, transfer learning, speech recognition, machine translation, and speech synthesis, Microsoft demonstrated this technology at the Code Conference, and posted videos of the demonstration on its blog on May 27, 2014: ## A Computer Masters Heads-up Limit Texas Hold 'em Poker January 8, 2015 A breakthrough in artificial intelligence published in January 2015 allowed a computer to master the simplest two-person version of the poker game known as Texas Hold'em working through every possible variation of play to make the perfect move every time. When performed without mistakes, just like the childhood game tic-tac-toe, there’s no way to lose. In this case the player is Cepheus, an algorithm designed by Canadian researchers. We have a strategy that can guarantee a player won’t lose,” said Michael Bowling, a computer scientist from the University of Alberta, who led a team working on the program. “It’s going to be a break-even game. It’s only when someone makes a mistake that they could end up losing. Michael BowlingNeil BurchMichael JohansonOskari Tammelin, "Heads-up limit hold'em poker is solved," Science 347, no. 6218 (2015) 145-149 "Poker is a family of games that exhibit imperfect information, where players do not have full knowledge of past events. Whereas many perfect-information games have been solved (e.g., Connect Four and checkers), no nontrivial imperfect-information game played competitively by humans has previously been solved. Here, we announce that heads-up limit Texas hold’em is now essentially weakly solved. Furthermore, this computation formally proves the common wisdom that the dealer in the game holds a substantial advantage. This result was enabled by a new algorithm, CFR, which is capable of solving extensive-form games orders of magnitude larger than previously possible" (Abstract). See also: http://news.sciencemag.org/math/2015/01/texas-hold-em-poker-solved-computer, accessed 01-14-2015. ## A Machine Vision Algorithm Learns to Attribute Paintings to Specific Artists May 2015 In May 2015 Babak Saleh and Ahmed Elgammal of the Department of Compuer Science, Rutgers University, described an algorithm that could recognize the Style, Genre, and Artist of a painting. "Saleh and Elgammal begin with a database of images of more than 80,000 paintings by more than a 1,000 artists spanning 15 centuries. These paintings cover 27 different styles, each with more than 1,500 examples. The researchers also classify the works by genre, such as interior, cityscape, landscape, and so on. "They then take a subset of the images and use them to train various kinds of state-of-the-art machine-learning algorithms to pick out certain features. These include general, low-level features such as the overall color, as well as more advanced features that describe the objects in the image, such as a horse and a cross. The end result is a vector-like description of each painting that contains 400 different dimensions. "The researchers then test the algorithm on a set of paintings it has not yet seen. And the results are impressive. Their new approach can accurately identify the artist in over 60 percent of the paintings it sees and identify the style in 45 percent of them. "But crucially, the machine-learning approach provides an insight into the nature of fine art that is otherwise hard even for humans to develop. This comes from analyzing the paintings that the algorithm finds difficult to classify. "For example, Saleh and Elgammal say their new approach finds it hard to distinguish between works painted by Camille Pissarro and Claude Monet. But a little research on these artists quickly reveals both were active in France in the late 19th and early 20th centuries and that both attended the Académie Suisse in Paris. An expert might also know that Pissarro and Monet were good friends and shared many experiences that informed their art. So the fact that their work is similar is no surprise. "As another example, the new approach confuses works by Claude Monet and the American impressionist Childe Hassam, who, it turns out, was strongly influenced by the French impressionists and Monet in particular. These are links that might take a human some time to discover" (MIT Technology Review May 11, 2015). Saleh, Babak, and Elgammal, Ahmed," Large-scale Classification of Fine-Art Paintings; Learning the Right Metric on the Right Feature" (http://arxiv.org/pdf/1505.00855v1.pdf, 5 May 2015. ## Based on a Single Example, A. I. Surpasses Human Capabilities in Reading and Copying Written Characters December 11, 2015 On December 11, 2015 Brenden M. Lake, Rusian Salakhutdinov, and Joshua B. Tenenbaum, from MIT and New York University reported advances in artificial intelligence that surprassed human capabilities in reading and copying written characters. The key advance was that the algorithm outperformed humans in identifying written characters based on a single example. Until this time, machine learning algorithms typically required tens or hundres of examples to perform with similar accuracy. Lake, Salakutdinov, Tenebaum, "Human-level concept learning through probabilistic program induction", Science, 11 December 2015, 350, no. 6266, 1332-1338. On December 14, 2015 the entire text of this extraordinary paper was freely available online. I quote the first 3 paragraphs: "Despite remarkable advances in artificial intelligence and machine learning, two aspects of human conceptual knowledge have eluded machine systems. First, for most interesting kinds of natural and man-made categories, people can learn a new concept from just one or a handful of examples, whereas standard algorithms in machine learning require tens or hundreds of examples to perform similarly. For instance, people may only need to see one example of a novel two-wheeled vehicle (Fig. 1A) in order to grasp the boundaries of the new concept, and even children can make meaningful generalizations via “one-shot learning” (13). In contrast, many of the leading approaches in machine learning are also the most data-hungry, especially “deep learning” models that have achieved new levels of performance on object and speech recognition benchmarks (49). Second, people learn richer representations than machines do, even for simple concepts (Fig. 1B), using them for a wider range of functions, including (Fig. 1, ii) creating new exemplars (10), (Fig. 1, iii) parsing objects into parts and relations (11), and (Fig. 1, iv) creating new abstract categories of objects based on existing categories (1213). In contrast, the best machine classifiers do not perform these additional functions, which are rarely studied and usually require specialized algorithms. A central challenge is to explain these two aspects of human-level concept learning: How do people learn new concepts from just one or a few examples? And how do people learn such abstract, rich, and flexible representations? An even greater challenge arises when putting them together: How can learning succeed from such sparse data yet also produce such rich representations? For any theory of learning (41416), fitting a more complicated model requires more data, not less, in order to achieve some measure of good generalization, usually the difference in performance between new and old examples. Nonetheless, people seem to navigate this trade-off with remarkable agility, learning rich concepts that generalize well from sparse data. "This paper introduces the Bayesian program learning (BPL) framework, capable of learning a large class of visual concepts from just a single example and generalizing in ways that are mostly indistinguishable from people. Concepts are represented as simple probabilistic programs— that is, probabilistic generative models expressed as structured procedures in an abstract description language (1718). Our framework brings together three key ideas—compositionality, causality, and learning to learn—that have been separately influential in cognitive science and machine learning over the past several decades (1922). As programs, rich concepts can be built “compositionally” from simpler primitives. Their probabilistic semantics handle noise and support creative generalizations in a procedural form that (unlike other probabilistic models) naturally captures the abstract “causal” structure of the real-world processes that produce examples of a category. Learning proceeds by constructing programs that best explain the observations under a Bayesian criterion, and the model “learns to learn” (2324) by developing hierarchical priors that allow previous experience with related concepts to ease learning of new concepts (2526). These priors represent a learned inductive bias (27) that abstracts the key regularities and dimensions of variation holding across both types of concepts and across instances (or tokens) of a concept in a given domain. In short, BPL can construct new programs by reusing the pieces of existing ones, capturing the causal and compositional properties of real-world generative processes operating on multiple scales. I"n addition to developing the approach sketched above, we directly compared people, BPL, and other computational approaches on a set of five challenging concept learning tasks (Fig. 1B). The tasks use simple visual concepts from Omniglot, a data set we collected of multiple examples of 1623 handwritten characters from 50 writing systems (Fig. 2)(see acknowledgments). Both images and pen strokes were collected (see below) as detailed in section S1 of the online supplementary materials. Handwritten characters are well suited for comparing human and machine learning on a relatively even footing: They are both cognitively natural and often used as a benchmark for comparing learning algorithms. Whereas machine learning algorithms are typically evaluated after hundreds or thousands of training examples per class (5), we evaluated the tasks of classification, parsing (Fig. 1B, iii), and generation (Fig. 1B, ii) of new examples in their most challenging form: after just one example of a new concept. We also investigated more creative tasks that asked people and computational models to generate new concepts (Fig. 1B, iv). BPL was compared with three deep learning models, a classic pattern recognition algorithm, and various lesioned versions of the model—a breadth of comparisons that serve to isolate the role of each modeling ingredient (see section S4 for descriptions of alternative models). We compare with two varieties of deep convolutional networks (28), representative of the current leading approaches to object recognition (7), and a hierarchical deep (HD) model (29), a probabilistic model needed for our more generative tasks and specialized for one-shot learning." ### 2016 – Present ## DeepMind's AI Algorithm Masters the Ancient Game of Go January 27, 2016 On January 27, 2016 the artificial intelligence company DeepMind, a division of Google based in London, announced that its AI progam AlphaGo mastered the ancient Chinese game of Go. "Traditional AI methods—which construct a search tree over all possible positions—don’t have a chance in Go. So when we set out to crack Go, we took a different approach. We built a system, AlphaGo, that combines an advanced tree search with deep neural networks. These neural networks take a description of the Go board as an input and process it through 12 different network layers containing millions of neuron-like connections. One neural network, the “policy network,” selects the next move to play. The other neural network, the “value network,” predicts the winner of the game. "We trained the neural networks on 30 million moves from games played by human experts, until it could predict the human move 57 percent of the time (the previous record before AlphaGo was 44 percent). But our goal is to beat the best human players, not just mimic them. To do this, AlphaGo learned to discover new strategies for itself, by playing thousands of games between its neural networks, and adjusting the connections using a trial-and-error process known as reinforcement learning. Of course, all of this requires a huge amount of computing power, so we made extensive use of Google Cloud Platform. "After all that training it was time to put AlphaGo to the test. First, we held a tournament between AlphaGo and the other top programs at the forefront of computer Go. AlphaGo won all but one of its 500 games against these programs. So the next step was to invite the reigning three-time European Go champion Fan Hui—an elite professional player who has devoted his life to Go since the age of 12—to our London office for a challenge match. In a closed-doors match last October, AlphaGo won by 5 games to 0. It was the first time a computer program has ever beaten a professional Go player. You can find out more in our paper, which was published in Nature today.... "We are thrilled to have mastered Go and thus achieved one of the grand challenges of AI. However, the most significant aspect of all this for us is that AlphaGo isn’t just an“expert” system built with hand-crafted rules; instead it uses general machine learning techniques to figure out for itself how to win at Go. While games are the perfect platform for developing and testing AI algorithms quickly and efficiently, ultimately we want to apply these techniques to important real-world problems. Because the methods we’ve used are general-purpose, our hope is that one day they could be extended to help us address some of society’s toughest and most pressing problems, from climate modelling to complex disease analysis. We’re excited to see what we can use this technology to tackle next!. Posted by Demis Hassabis, Google DeepMind" (https://googleblog.blogspot.com/2016/01/alphago-machine-learning-game-go.html, accessed 02-09-2016). ## What Google's DeepMind Learned in Seoul with AlphaGo March 16, 2016 ## What we learned in Seoul with AlphaGo Go isn’t just a game—it’s a living, breathing culture of players, analysts, fans, and legends. Over the last 10 days in Seoul, South Korea, we’ve been lucky enough to witness some of that incredible excitement firsthand. We've also had the chance to see something that's never happened before: DeepMind's AlphaGo took on and defeated legendary Go player, Lee Sedol (9-dan professional with 18 world titles), marking a major milestone for artificial intelligence. Pedestrians checking in on the AlphaGo vs. Lee Sedol Go match on the streets of Seoul (March 13) Go may be one of the oldest games in existence, but the attention to our five-game tournament exceeded even our wildest imaginations. Searches for Go rules and Go boards spiked in the U.S. In China, tens of millions watched live streams of the matches, and the “Man vs. Machine Go Showdown” hashtag saw 200 million pageviews on Sina Weibo. Sales of Go boards even surged in Korea. Our public test of AlphaGo, however, was about more than winning at Go. We founded DeepMind in 2010 to create general-purpose artificial intelligence (AI) that can learn on its own—and, eventually, be used as a tool to help society solve some of its biggest and most pressing problems, from climate change to disease diagnosis. Like many researchers before us, we've been developing and testing our algorithms through games. We first revealed AlphaGo in January—the first AI program that could beat a professional player at the most complex board game mankind has devised, using deep learning and reinforcement learning. The ultimate challenge was for AlphaGo to take on the best Go player of the past decade—Lee Sedol. To everyone's surprise, including ours, AlphaGo won four of the five games. Commentators noted that AlphaGo played many unprecedented, creative, and even“beautiful” moves. Based on our data, AlphaGo’s bold move 37 in Game 2 had a 1 in 10,000 chance of being played by a human. Lee countered with innovative moves of his own, such as his move 78 against AlphaGo in Game 4—again, a 1 in 10,000 chance of being played—which ultimately resulted in a win. The final score was 4-1. We're contributing the$1 million in prize money to organizations that support science, technology, engineering and math (STEM) education and Go, as well as UNICEF.

We’ve learned two important things from this experience. First, this test bodes well for AI’s potential in solving other problems. AlphaGo has the ability to look “globally” across a board—and find solutions that humans either have been trained not to play or would not consider. This has huge potential for using AlphaGo-like technology to find solutions that humans don’t necessarily see in other areas. Second, while the match has been widely billed as "man vs. machine," AlphaGo is really a human achievement. Lee Sedol and the AlphaGo team both pushed each other toward new ideas, opportunities and solutions—and in the long run that's something we all stand to benefit from.

But as they say about Go in Korean: “Don’t be arrogant when you win or you’ll lose your luck.” This is just one small, albeit significant, step along the way to making machines smart. We’ve demonstrated that our cutting edge deep reinforcement learning techniques can be used to make strong Go and Atari players. Deep neural networks are already used at Google for specific tasks—like image recognitionspeech recognition, and Search ranking. However, we’re still a long way from a machine that can learn to flexibly perform the full range of intellectual tasks a human can—the hallmark of trueartificial general intelligence.
Demis and Lee Sedol hold up the signed Go board from the Google DeepMind Challenge Match

With this tournament, we wanted to test the limits of AlphaGo. The genius of Lee Sedol did that brilliantly—and we’ll spend the next few weeks studying the games he and AlphaGo played in detail. And because the machine learning methods we’ve used in AlphaGo are general purpose, we hope to apply some of these techniques to other challenges in the future. Game on!