Title: Minimalistic Neural Models for Cognitive Algorithms, or an Introduction to NEMO
Abstract:
In this talk I will introduce NEMO, which builds on the previous "assembly calculus" model. We hope to develop a "computer science of the brain" by defining computational models using abstractions of the most basic neurobiological principles, with the goal of finding a minimal set of assumptions that could model cognition. We'll define a model using only spiking neurons, random synapses, local inhibition, and local plasticity. We’ll show that when we define these ingredients in a very simple way, we obtain a model where interconnected co-firing sets of neurons, or assemblies, emerge from the dynamical system. It turns out it is possible to carry out complex operations on these assemblies, such as copying and merging them. This suggests that a much richer computational system may be possible in this unusual, but biologically motivated, model of computation.
I will then present the "highlights" of NEMO of the past few years-- the most interesting algorithms that can be implemented in NEMO that could model cognitive phenomena. These include an efficient parser (of context free languages), a "baby brain" model that can learn the meaning of words and basic word order (syntax) from grounded sentences, and an algorithm for learning and sampling from a discrete conditional distribution.