Weill Cornell Medicine Events

Department of Physiology and Biophysics Special Seminar


Kanaka Rajan, Ph.D.

Biophysics Theory Fellow

Lewis-Sigler Institute for Integrative Genomics

Princeton University



“Recurrent Network Models of Sequence Generation and Memory”


Abstract: The sequential activation of neurons is a common feature of network activity during a variety of behaviors and has been proposed as a mechanism for cortical computation, including short term memory. Previous modeling approaches for sequences and memory networks have emphasized highly specialized architectures in which a principled mechanism is pre-wired into the connectivity of the network. In this work, we demonstrate that starting from random synaptic connectivity and allowing a small fraction of connections to undergo modification, a largely disordered recurrent network can produce sequences and short-term memory. We use this process, which we call Partial In-Network training (PINning), to model and match data from cellular-resolution imaging of neural activity in the mouse posterior parietal cortex (PPC) during a memory-guided two-alternative forced choice task in a virtual environment [Harvey, Coen & Tank, 2012]. In the model, as in the PPC data, individual neurons exhibit transient activations that are staggered relative to one another in time to form sequences spanning the duration of the task, and different sequences are activated on trials with different cues and choices. Analysis of the connectivity matrices of the minimally structured model networks revealed that the time-ordered neural activity is produced by the cooperation between recurrent synaptic interactions and external inputs, rather than feedforward connections, or the asymmetric connections of ring attractor models for sequences. In addition, our model showed that sequential activation across a population of neurons is an efficient mechanism for implementing short-term memory with comparable memory capabilities to previously proposed fixed point mechanisms. Together our results [Rajan, Harvey & Tank, 2016] develop a modeling framework based on generic, flexible, and minimally modified networks and suggest that neural activity sequences may emerge through learning from largely unstructured network architectures. 


Emre Aksay, Ph.D.

Wednesday, June 21, 2017 at 11:00am to 12:00pm

Weill Cornell Medical College, LC-504 Conference Room
1300 York Avenue, New York NY

Recent Activity