In January 1952 Marvin Minsky, a graduate student at Harvard University Psychological Laboratories implemented the SNARC (Stochastic Neural Analog Reinforcement Calculator). This randomly connected network of Hebb synapses was the first connectionist neural network learning machine that when "rewarded" facilitated recently-used pathways. The SNARC, implemented using vacuum tubes, was possibly the first artificial self-learning machine.
Minsky, A Neural-Analogue Calculator Based upon a Probability Model of Reinforcement," Harvard University Psychological Laboratories, Cambridge, Massachusetts, January 8, 1952. This reference came from Minsky's bibliography of his selected publications on his website in December 2013. He did not include this in his bibliography on AI in Computers and Thought (1963), leading me to believe that some or all of the information may have been included in his Princeton Ph.D. dissertation, Neural Nets and the Brain Model Problem (1954). That was also unpublished.