Memory reactivation and the apparent biological implausibility of CLST
Professor Bruce L McNaughton, University of California at Irvine, USA
Two issues with CLST are: 1) consolidation of new memories appears to require many cycles of reactivation of new data interleaved with all previously acquired experience. Constraints on available reactivation time probably renders this unrealistic. 2) connectionist models avoid CI by retraining on all previous data, but the brain only has access to stored representations.
McClelland et al.'s (1995), 'catastrophic' introduction of a penguin into the network, without interleaved retraining, was less than completely catastrophic: almost all the error was in the animals. The plant category was hardly affected. This leads to the hypothesis that exhaustive reactivation is not always necessary, and can be substituted with 'Similarity Weighted Interleaved Learning (SWIL)' in which only stored items that are similar to new items (e.g. the other animals) need to be interleaved in the reactivation mix. Under what circumstances this does or does not work will be explored in Jay's talk. I propose a simple, attractor style, hypothesis about how SWIL might occur in hippocampal-cortical interactions. A possible solution to the second problem was proposed in 1995 by Robins, with his concept of pseudorehearsal, in which random patterns were interleaved with new data (https://arxiv.org/abs/1802.03875). SWIL could operate on a similar principle, plus a similarity weighting based on experience-dependent suppression of AHPs (Disterhoft TINS, 2006, 29:587), which could bias recently partially activated cortical attractors to reactivate spontaneously when "pseudopatterns" (i.e. random inputs) are presented (Shen, Hippocampus, 1996, 6:685). By definition, these would be stored patterns that overlap with the new input.
Integration of new information in memory: new insights from a complementary learning systems perspective
Dr James L McClelland, Stanford University, USA
According to complementary learning systems theory (CLST), integrating new memories into a brain-like neural networks without interfering with what is already known depends on a gradual learning process, interleaving new items with items previously learned. However, empirical findings now establish that information consistent with prior knowledge can sometimes be integrated quickly with little interference, and recent modeling research indicates that this finding can be captured in neural network models that reflect the properties of the neocortical learning system proposed in CLST. New work in collaboration with Bruce McNaughton and Andrew Lampinen uses deep linear neural networks in hierarchically structured environments to gain new insights into when integration is fast or slow and how integration might be made more efficient. The environments correspond to familiar taxonomic hierarchies, where items separated low in the tree (e.g., different species of sea gulls) share nearly all their properties, but items that separate at higher branches (sea gulls vs pine trees) share far fewer properties. Deep linear networks learn this kind of domain structure in a gradual, stage-like progression, capturing successive split in the hierarchy after increasingly long delays. In this context, a new item to can be characterized in terms of its projection onto the known hierarchy and whether it adds a new categorical split. The projection onto the known hierarchy can be learned rapidly without interleaving, but if the item has unique features or feature combinations requiring a new split, integration will require gradual interleaved learning. When the new item only overlaps with items in a branch of the hierarchy, interleaving can be focused on these items, with less interleaving overall. Discussion will consider how the brain might exploit these facts to make learning more efficient and will highlight predictions about what aspects of new information might be hard or easy to learn.