Dynamic probabilistic inference in the brain

Organizers

Anna Kutschireiter | Harvard Medical School, Cambridge, USA
Jan Drugowitsch | Harvard Medical School, Cambridge, USA

Abstract

Every day, our brain needs to make sense of the rich, dynamical stream of sensory
inputs, and combine it with prior knowledge about its environment. Ample behavioral
evidence suggests that the brain’s processing of information conforms to the rules of
probabilistic inference. Most of this evidence came from static trial-by-trial experiments
that do not reflect the dynamic nature of our environment, leading to simplified and
rather restricted models of how our brains perform such inference. The aim of this
workshop is to look beyond such simplified, static models of inference, and ask how the
brain could perform the continuous-time dynamic inference required to operate in
natural environments. Such inference needs to span the range of synapses learning
environmental regularities, over the efficient and effective processing of dynamic and
continuously changing sensory inputs, to applying continuous-time control in order to
act upon the world’s inferred state. Recent experiments have started moving towards
more natural behaviors and as such provide the ideal benchmark to test the emerging
models against.

We will bring together researchers working on models of dynamical inference in the brain,
ranging from inference and learning on the level of synapses, single neurons and neuronal
networks to predictions of optimal strategies and behavior, as well as on experiments to test these predictions. More precisely, the goal is to provide a forum to discuss recent developments on all these levels and consider the implications of adding a dynamic component to the usually static inference. By spanning a wide range of research areas, the workshop should appeal to the broad audience attending the main conference.

Workshop schedule
session 1: choice of representation
session 2: synapse level. learning
session 3: behaviour & decision making
session 4: perceptual inference & efficient coding

Schedule

time (CEST)
14:00	Introduction
14:15	Cristina Savin \| New York University, USA Neural representations of uncertainty: the right tool for the right job
14:45	Eszter Vértes \| DeepMind, London, UK Distributed distributional codes for learning successor features in partially observable environments
15:15	Máté Lengyel \| Cambridge University, UK Cortical-like dynamics in recurrent circuits optimized for sampling-based probabilistic inference
15:45	30 min break
16:15	Robert Legenstein \| Graz University of Technology, Austria Inference from Facts with Hebbian Plasticity
16:45	Jannes Jegminat \| University of Bern, Switzerland Learning as filtering: a framework for spike based plasticity
17:15	Joseph Makin \| Purdue University, West Lafayette, Indiana, USA Sensory Integration, Density Estimation, and Information Retention
17:45	60 min break
18:45	Dimitrije Marković \| Technical University Dresden, Germany Anticipating changes: decision-making with temporal expectations
19:15	Zachary Kilpatrick \| University of Colorado Boulder, USA Dynamic forms of urgency in ideal and human observers
19:45	Anja Zai \| ETH Zuerich, Switzerland Inferring the trial-by-trial structure of reinforcement learning with latent variable models
20:15	15 min break
20:30	Ralf Haefner \| University of Rochester, USA A perceptual confirmation bias as the result of feedback dynamics arising during approximate hierarchical inference
21:00	Ann Hermundstad \| HHMI Janelia Research Campus, Ashburn, USA Adaptability and efficiency in neural coding
21:30	Final discussion
22:00	End