Learning Machine
A learning machine is a machine whose behavior improves with experience rather than being fully fixed by its designer in advance. The idea took shape around 1948 to 1959, when several researchers independently argued that the practical route to machine intelligence might be to build something that could be taught or could adapt, instead of trying to write down adult-level competence directly. Behind it lies a change in what a program is: not a finished list of instructions but an initial state plus a procedure for changing that state in response to feedback.
Turing gave the idea its most influential statement. In his 1948 National Physical Laboratory report Intelligent Machinery, not published until 1968, he described networks of "unorganised machines" that could be modified by interference or training, and in section 7 of Computing Machinery and Intelligence (Mind, 1950) he proposed building a child machine and educating it instead of programming an adult mind. Education would involve rewards and punishments, with the teacher unable to foresee every detail of what the machine ended up doing; he also noted that such a machine could surprise its makers. The proposal rests on a cost argument: the initial program might be simpler than the finished mind, and the hard-to-specify content could be absorbed from experience.
While Turing wrote, other workers built physical machines that adapted. W. Ross Ashby's Homeostat, announced in 1948 and later analyzed in his book Design for a Brain (1952), was an electromechanical device of four interconnected units that, when disturbed beyond a limit, randomly rearranged its own parameters until it found a stable configuration; Ashby called the property ultrastability. W. Grey Walter in Bristol built the tortoises Elsie and Elmer in 1948 and 1949, small wheeled robots with a light sensor, touch sensor and two vacuum-tube amplifiers that seemed to seek light, avoid obstacles and return to a charging hutch. The first tortoises were reactive rather than learning, but later versions added a conditioning circuit, and the project showed that lifelike behavior needs little machinery.
Arthur Samuel turned the idea into the first widely known learning program. Working at IBM from the early 1950s, he wrote a checkers player that could improve by playing; his 1959 paper "Some Studies in Machine Learning Using the Game of Checkers" in the IBM Journal of Research and Development described two methods, rote learning, which stored positions with their evaluated values, and adjusting the coefficients of a scoring polynomial through self-play. The program eventually played better than its author, which was striking given the 1950s hardware, and the paper is the usual source for the phrase machine learning. Samuel's evaluation-by-lookahead method also foreshadowed later temporal-difference learning.
Frank Rosenblatt supplied the neural network branch. At the Cornell Aeronautical Laboratory he reported the perceptron in 1957 and published it in Psychological Review in 1958, with a Mark I Perceptron built in hardware to recognize simple images, its weights adjusted by motors whenever it made an error. The perceptron convergence theorem showed that for linearly separable data the learning rule would find a separating solution in finitely many steps. The enthusiasm and later disappointment around that work is a separate story told in the entry on the perceptron.
The unifying structure is the feedback loop, and the main historical dispute was about where the knowledge should live. In Turing's picture the knowledge enters through education, in Ashby's through random search constrained by stability, in Samuel's through self-play, and in Rosenblatt's through labeled examples. Critics pointed out that learning alone does not remove the need for structure: a learner can only adjust what its representation allows, which is why the symbolic school of the 1960s and 1970s put its effort into knowledge representation rather than adaptation, and why expert systems relied on hand-encoded rules. Machine learning returned to dominance when data and computing power made flexible learners practical.
Several misconceptions are worth correcting. A learning machine does not necessarily become intelligent through learning; Samuel's program learned a narrow skill and Grey Walter's tortoises were not learning in their simplest form. Learning does not mean the machine is free of human design, since the architecture, the feedback signal and the training data embody many human choices. And it is not a modern invention: the vocabulary of training, rewards and generalization was in place in the 1950s, and modern systems differ mainly in scale, in data, and in methods such as backpropagation, not in the basic loop.
Read learning machine as the shift from writing down what a machine should do to arranging for it to find out: design the learner and the feedback, then let experience supply the content.