All animals show some form of learning, but the range, sublety and pervasiveness of human learning is the main thing that sets human beings apart from other species. How does the nervous system work to produce learning? There are two ways to approach this problem: bottom up, and top down. The bottom-up approach begins with the properties of individual neurons and synapses and aims to understand complex learning through exhaustive analysis of simple neural circuits and simple learning processes. This approach is being applied very successfully to elementary learning in several invertebrate species, but it is still a long way from the complexities of learning in higher animals. A complementary alternative is the top down approach, which underlies the work here proposed. The top-down method is to begin with careful behavioral analysis of apparently complex tasks that can nevertheless be reduced to simple performance rules. Theoretical analysis can often suggest formal real-time models that behave in the ways described by a given performance rule. A dynamic model of this sort that survives rigorous behavioral testing is likely to reflect enduring and measurable properties of the underlying neural machinery. We have found a very simple class of performance rules, the value-transfer hypothesis, that may underlie animals' ability to solve a """"""""reasoning"""""""" task, transitive inference. We have also proposed a very simple dynamic model for the assignment-of-credit problem in operant conditioning, that is, the process by which a response is selected by consequential reinforcement. And most recently we have discovered a simple way to produce sequence learning in a recurrent """"""""neural"""""""" network. The proposal describes additional experimental tests of the value-transfer hypothesis, and theoretical explorations of learning systems built out of the assignment- of-credit model and different forms of neural network. Our immediate objective is to arrive at a dynamic model for the transitive-inference task and for a range of other discrimination-learning tasks. Our ultimate objective is to use these models as guides to understanding the role of the nervous system in learning.
