Gatsby Computational Neuroscience Unit

UCL Gatsby Unit

Mate Lengyel

Wednesday 8th May 2013

Time: 4pm

Basement Seminar Room

Alexandra House, 17 Queen Square, London, WC1N 3AR

Biography

Máté Lengyel is a Reader in Computational Neuroscience at the Department of Engineering, University of Cambridge. Previously, he studied biology for his MSc and neurobiology for his PhD at Eötvös University, Budapest. He was then a postdoc with Peter Dayan at the Gatsby Computational Neuroscience Unit, UCL, followed by a visiting research fellowship at the Collegium Budapest Institute for Advanced Study. He studies learning and memory from computational, algorithmic/representational and neurobiological viewpoints.

Probabilistic inference in neural circuits: taming monsters and other tales

I will talk about two different levels at which neural circuits need to perform probabilistic inference.

First is the level of a single cell that needs to compute some function of its presynaptic partners' activities and represent the outcome of this computation in its own activity. While such computations are most easily, and certainly most commonly, formalised as transformations of continuous variables, neurons communicate with digital signals: spikes. However, spikes only give a limited amount of information about the underlying activities of presynaptic neurons. Thus, the postsynaptic neuron needs to infer what its output should be based on the incoming spikes of its presynaptic partners. We show, both numerically and analytically, that the form of this inference closely resembles the operations that active dendrites perform on their inputs, which suggests a normative role for such dendritic nonlinearities, and a matching of systems level-properties of neural circuits (joint activity distributions of presynaptic populations), and cellular-level properties of single neurons.

Second is the level of cortical populations that need to infer environmental features form partial, ambiguous, and noisy sensory input. We have put forward a hypothesis, the sampling hypothesis, that relates directly the variability in neural responses to the uncertainty captured by probability distributions over environmental features that the cortex represents. I will revisit some results we published earlier in support of this hypothesis, and that have been recently challenged in particular regarding the role of learning. I will show how new maximum entropy models can and should be used to assess the role of pairwise correlations in shaping spontaneous and evoked cortical activities, and their similarity. I will also describe pilot results addressing the role of experience in the development of this similarity. I will also briefly mention some work under way in which we are testing psychophysical predictions of the sampling hypothesis.