— Ch. 1 · The Goof Button And The First Learner —
Machine learning.
~7 min read · Ch. 1 of 7
In 1959, Arthur Samuel coined the term machine learning while working at IBM. He created a computer program that could calculate winning chances in checkers for each side. This early system marked the first time a computer learned from experience rather than following fixed instructions. A representative book on research into machine learning during the 1960s was Nils Nilsson's Learning Machines. It dealt mostly with pattern classification and set the stage for future developments.
By the early 1960s, Raytheon Company developed an experimental learning machine called Cybertron. This device used punched tape memory to analyze sonar signals, electrocardiograms, and speech patterns. Operators taught it rudimentary reinforcement learning techniques through repetitive training sessions. The system included a "goof button" that allowed human teachers to force reevaluation of incorrect decisions. Time magazine reported on this mechanism in the 18th of August 1961 under the headline Science: The Goof Button.
Interest in pattern recognition continued into the 1970s as described by Duda and Hart in their 1973 publication Pattern Recognition and Scene Analysis. In 1981, researchers presented a report on teaching strategies where artificial neural networks learned to recognize 40 characters from a computer terminal. These characters included 26 letters, 10 digits, and four special symbols. Tom M. Mitchell later provided a widely quoted definition stating that a program learns if its performance improves with experience.
Three Paths To Knowledge
Modern-day machine learning algorithms break into three distinct types based on how they receive feedback. Supervised learning presents computers with example inputs and desired outputs given by a teacher. The goal is to learn a general rule mapping inputs to outputs. Classification algorithms handle limited output values while regression algorithms manage any numerical value within a range. Email filtering serves as an example where input is an incoming message and output determines which folder to file it.
Unsupervised learning operates without labels or explicit guidance. Algorithms identify commonalities in data and react based on presence or absence of such patterns. Central applications include clustering, dimensionality reduction, and density estimation. Cluster analysis assigns observations into subsets so members within the same cluster are similar according to predesignated criteria. Principal component analysis transforms higher-dimensional data like three dimensions down to smaller spaces such as two dimensions.
Reinforcement learning focuses on decisions made regarding unknown time periods through interaction with dynamic environments. Programs navigate problem spaces receiving feedback analogous to rewards they try to maximize. Autonomous vehicles and game-playing systems against human opponents utilize these techniques. Many reinforcement learning algorithms use dynamic programming methods when exact mathematical models prove infeasible. The environment typically represents a Markov decision process allowing agents to take actions that produce cumulative rewards.