Daniel Soudry: Jan 21st

Daniel Soudry will talk about the following paper:

Title: Fixed-form variational posterior approximation through stochastic linear regression

Authors:Tim Salimans and David A. Knowles

Abstract: We propose a general algorithm for approximating nonstandard Bayesian posterior distributions. The algorithm minimizes the Kullback-Leibler divergence of an approximating distribution to the intractable posterior distribution. Our method can be used to approximate any posterior distribution, provided that it is given in closed form up to the proportionality constant. The approximation can be any distribution in the exponential family or any mixture of such distributions, which means that it can be made arbitrarily precise. Several examples illustrate the speed and accuracy of our approximation method in practice.

Moritz Deger: Dec 17th

Dynamics and estimation of large-scale spiking neuronal network models

Computations in neural circuits emerge from the interaction of many neurons. Although accurate single neuron models exist, little is known about the synaptic connectivity of large-scale circuits of neurons. However, recent experimental techniques enable recordings of the activity of thousands of neurons in parallel. Such data hold the promise that inferences about the underlying synaptic connectivity can be made. I will present two complementary approaches to gain insights into the collective dynamics of neural circuits.
On the one hand, I will report on recent progress in the reconstruction of networks of 1000 simulated spiking neurons by maximum likelihood estimation of a generalized linear model (GLM), in which a million possible synaptic efficacies have to be determined. The work demonstrates that reconstructing the connectivity of thousands of neurons is feasible, and that hidden embedded subpopulations can be detected in the reconstructed connectivity.

On the other hand, spiking GLM's are a versatile class of single neuron models that can represent important features such as intrinsic stochasticity, refractoriness, and spike-frequency adaptation. For this class of models, I will present a new theory of population dynamics, which takes into account finite-size fluctuations and accurately describes the population-averaged neural activity in time, in response to arbitrary stimuli. Based on this theory, GLM network models with parameters extracted from data can be used to understand neural processing in realistic networks, and circuit-level information processing can be explained.

Post-NIPS special discussion: Dec 16th 2-3:30pm

Ferran Diego: Dec 3rd

Identifying Neuronal Activity  from Calcium Imaging Sequences

Calcium imaging is an increasingly popular technique for monitoring simultaneously the neuronal activity of hundreds of cells at single cell resolution. This makes an essential tool for studying spatio-temporal patterns of distributed activity that are crucial determinants of behavioral and cognitive functions such as perception, memory formation, motor activity, decision making and emotion. However, most approaches only focus on identifying positions of each cell (or parts of cells) by eye or semi-automated. Therefore, in this talk, we present two main approaches for detecting automatically neural activity. The first approach formulates the identification of neuronal activity of single cells and the neuronal co-activation into the same framework. That is, we propose a decomposition of the matrix of observations into a product of more than two sparse matrices, with the rank decreasing from lower to higher levels, driven by the semantic concepts of pixel -> neuron -> assembly in the neurosciences image sequence. In contrast to prior work, we allow for both hierarchical and heterarchical relations of lower-level to higher-level concepts.
Moreover, the proposed bilevel SHMF (sparse heterarchical matrix factorization) is the first formalism that allows to simultaneously interpret a calcium imaging sequence in terms of the constituent neurons, their membership in assemblies, and the time courses of both neurons and assemblies.  The second approach describes a unified formulation and algorithm to find an extremely sparse representation for Calcium image sequences in terms of cell locations, cell shapes, spike timings and impulse responses. Solution of a single optimization problem yields cell segmentations and activity estimates that are on par with the state of the art, without the need for heuristic pre- or postprocessing."

Franck Polleux: Nov 19th

The talk was canceled

Ethan S. Bromberg-Martin: Nov 12th

What does information seeking tell us about reinforcement learning?

Conventional theories of reinforcement learning explain how we choose actions to gain rewards, but we also often choose actions to help us predict rewards. This behavior is known as information seeking (or 'early resolution of uncertainty') in economics and a form of ‘observing behavior’ in psychology, and is found in both humans and animals. We recently showed that the preference to gather information about future rewards is signaled by many of the same neurons that signal preferences for appetitive rewards like food and water. This suggests that information seeking and conventional reward seeking share a common neural mechanism.
At the moment, we know very little about the nature of these neural computations. A major roadblock is theoretical: most prominent theories of reinforcement learning were originally designed to account for appetitive reward seeking and are unable to account for information seeking. How can we address this gap in our theories? I will summarize the state of the field, including my own work and that of others, and use this to propose ways that we can revise current theories of reinforcement learning to account for information seeking.

Sundeep Rangan: November 5th

Approximate Message Passing for Inference in Generalized Linear Models

Generalized approximate message passing (GAMP) methods are a powerful new class of inference algorithms designed for generalized linear models, where an input vector x must be estimated from a noisy, possibly nonlinear function of a transform z = Ax.  The methods are based on Gaussian approximations of loopy BP and have the benefit of being computationally extremely simple and general. Moreover, under certain large random transforms, the algorithms are provably Bayes optimal, even in many non-convex problem instances. In this talk, I will provide an overview of GAMP methods, some of the recent extensions to unknown priors and structured uncertainty. I will also highlight some of the main issues in convergence of the algorithm and discuss some applications in neural connectivity detection from calcium imaging.

Bio:
Dr. Rangan received the B.A.Sc. at the University of Waterloo, Canada and the M.Sc. and Ph.D. at the University of California, Berkeley, all in Electrical Engineering.  He has held postdoctoral appointments at the University of Michigan, Ann Arbor and Bell Labs.  In 2000, he co-founded (with four others) Flarion Technologies, a spin off of Bell Labs, that developed Flash OFDM, the first cellular OFDM data system and pre-cursor to many 4G wireless technologies.  In 2006, Flarion was acquired by Qualcomm Technologies.  Dr. Rangan was a Director of Engineering at Qualcomm involved in OFDM infrastructure products.  He joined the ECE department at the NYU Polytechnic School of Engineering in 2010.  He is an IEEE Distinguished Lecturer of Vehicular Technology Society. His research interests are in wireless communications, signal processing, information theory and control theory.