Lateral Inhibition Facilitates Sequential Learning in a Hippocampus-Inspired Auto-Associator
Benjamin Midler, James McClelland, Stanford University, United States
Contributed Talks 4 Lecture
Grand Ballroom A-C
Sun, 28 Aug, 12:45 - 13:05 Pacific Time (UTC -8)
Functional and neurobiological characterizations of the hippocampus and its sub-structures have yielded improved theoretical understandings of how the brain implements learning and memory. Despite these advancements, much about how the architecture of these neural structures implement key memory abilities, such as resistance to catastrophic interference, remains to be explored. The present objective is to construct a neural network model imbued with key features of the hippocampus, such as forming latent representations and lateral inhibition, with the hypothesis that resistance to catastrophic interference will be an emergent property of this architecture. We demonstrate that such a model resists catastrophic interference on-par with established benchmarks and incentivizes unique solutions to the computational challenge of simultaneously maintaining new and existing memories.