Inductive Learning with External Representations

External representation is the use of the physical world for cognitive ends, the enlargement of the mechanisms of representation to include the actionperception cycle. It has recently been observed that such representation is pervasive in human activity in both pragmatic and more abstract tasks. It is argued here that by forcing an artificial learning system to off-load all of its representation onto a (simulated) external world, we may obtain a model that is biased in a very natural way to represent functional relations in ways similar to those used by people. After learning a function from examples, such a model should therefore generalize to unseen instances in ways that we would consider correct. These ideas are tested by developing two machine learning systems, in which representation relies on the sensorimotor control of simulated robotic agents. These systems are able to represent a variety of functional relations by means of their action and perception, and they learn to spontaneously do so from examples. Moreover, they generalize extremely well to unseen problems even after a small number of examples, including on functions such as n -parity that are notoriously difficult to generalize for machine learning algorithms. It is argued that despite these systems’ simplicity, the external representations that they evolve are similar to those used by people on