Spike-Based Bayesian Learning in Neocortical Microcircuits
CBN (Computational Biology and Neurocomputing) seminars
Tuesday 17 June 2014
to 11:00 at
Phil Tully (CB/CSC/KTH and Institute for Adaptive and Neural Computation, University of Edinburgh)
Large-scale, recurrently connected cortical circuits exhibit complex dynamical interactions, and play host to many plastic mechanisms that can sculpt and be sculpted by ongoing activity. But how can we begin to understand these intricate synergies in a principled way? We propose that the connectivity of a biophysical attractor memory network could be learned using the spike-based Bayesian Confidence Propagation Neural Network (BCPNN) learning rule. Although the approach encompasses a diversity of mechanisms including Hebbian, homeostatic synaptic, intrinsic, and neuromodulated plasticity, it is straightforwardly understood since it is neatly encapsulated within the framework of probabilistic inference .
In this talk I will focus on spike-based BCPNN learning using different time scales. I'll show how fast AMPA connections provide the recurrent excitation necessary for assembling neurons into stable groups, i.e. attractors, while slowly decaying NMDA receptor (NMDAR) conductances provide prolonged activations that act as bridges for connecting different attractors. Thus, NMDAR allows for the passage of representational content from one ensemble to the next in sequence, and propels the network along a trajectory through attractor state space. The resulting spatiotemporal activity patterns consist of intermittent population bursts with abrupt sequential transitions occurring on the order of hundreds of milliseconds, resembling dynamics widely observed across motor , sensory , memory  and decision-making  tasks. Overall, our work implies that the presence of a spike, or lack thereof, not only enacts measurable changes in the biochemical makeup of synapses and cells, but moreover contributes to an underlying, ongoing probabilistic inference process. We provide a biophysical realization of Bayes' Rule by reconciling several observed neural phenomena whose functional effects are only partially understood in concert.
1. Synaptic and nonsynaptic plasticity approximating probabilistic inference. Tully, Hennig & Lansner, Frontiers in Synaptic Neuroscience 6:8, 2014.
2. Parallel processing of serial movements in prefrontal cortex. Averbeck et. al., Proc. Natl. Acad. Sci. USA 99:20, 13172-13177, 2002.
3. Natural stimuli evoke dynamic sequences of states in sensory cortical ensembles. Jones et. al., Proc. Natl. Acad. Sci. USA 104:47, 18772-18777, 2007.
4. Behavior-dependent short-term assembly dynamics in the medial prefrontal cortex. Fujisawa et. al., Nature Neuroscience 11:7, 2008.
5. Successful choice behavior in associated with distinct and coherent network states in anterior cingulate cortex. Lapish et. al., Proc. Natl. Acad. Sci. USA, 105:33, 2008.