Document Actions

You are here: Home / Bernstein Conference / 2019 / Satellite Workshops / Neural computation through recurrent dynamics: from theory to experiment and back

Neural computation through recurrent dynamics: from theory to experiment and back


Dynamical systems theory has long provided a useful language to understand neural computations.  Recently, there is increasing support for applying this framework to understand how populations of neurons  work in concert to perform computations. Recurrent networks constitute an important class of models that generate rich dynamics while also including anatomically-relevant constraints. Further, they have proven to be of remarkable value in the interpretation and understanding of recent neural data. These models are  being used in three ways in neuroscience. First, to understand how behaviorally-relevant computations are  carried out in populations of recurrently connected neurons. Second, to develop theoretical frameworks that  provide a quantitative handle on why specific dynamics lead to task-relevant computations. Third, they are being used as tools to infer the underlying dynamics observed in data to refine our understanding of neural computations.

These three approaches to recurrent dynamical models have different goals. The goal of the first is to  understand why certain patterns of dynamics emerged in neural data while that of the second is to understand the mathematics underlying these dynamical computations. The goal of the third is to build tools, using the assumption of underlying recurrent dynamics, to infer characteristics of neural data. Each method  tackles the problem from a different vantage point but what can these approaches learn from each other?

There is much to be gained from a unified understanding of recurrent computation in the brain and this workshop aims to encourage exchanges of insights among these three perspectives. To this end, we bring together six researchers with recent innovative contributions to recurrent dynamics in theory, experimentation or application-related models. Structured discussions will help identify the disparities, the constraints and the  insights that can be shared among these three approaches to recurrent models in neuroscience.