The discovery of structure by multi-stream networks of local processors with contextual guidance

Abstract

We study multi-stream networks in which feature discovery and associative learning interact cooperatively at the level of the local processors. These processors select and recode the information in their receptive field (RF) inputs that is predictably related to the context within which it occurs. To enable them to do this they are provided with local contextual input in addition to their receptive field input. This input guides both learning and processing to the RF information that is related to the context, but without confounding the information that the processor transmits about the RF. We show that these nets can discover linear functions of their inputs that are predictably related across streams. They can do so whether or not these variables are the most informative within streams, and when there is no evidence within streams as to the existence of these variables. They discover the relevant variables concurrently with, and because of, discovering the predictive relations between them. Two-stage multi-stream nets can discover the XOR function when the only evidence for it is in the higher-order cross-stream statistical dependencies, and when there are no pair-wise correlations in the input either within or across streams. This problem is at the limits of the ability of these nets, however, and they do not always solve it. Various aspects of these limitations are studied. Overall the results show that performance improves as the number of streams increases, and this encourages the view that the approach may have biological relevance.