The Dynamics of Recurrent Neural Networks: Traveling Waves, Short Term Memory, and Rapid Reconfiguration of Neural Circuits
Event Details
- Type
- Monday Lecture Series
- Speaker(s)
-
Marcelo Magnasco, Ph.D., professor and head, Laboratory of Integrative Neuroscience, The Rockefeller University
- Speaker bio(s)
-
Dynamic reconfiguration is prominent in cortex. During any one task or behavioral state, the dynamics of neural assemblies appears to be fixed and operate on some low-dimensional “neural manifold”; upon changing the task rapid switches are observed, from changes in timescales and lengthscales to changing the manifold itself. Cortex thus appears to be a machine for creating a huge number of ephemeral dynamical circuits.
This has caused a rift between the "circuit view" and the "effective dynamics view", whose resolution has been sought a search for the gating circuitry: ad-hoc circuit elements that shunt communication. While gating circuits do exist (in the spinal cord and brain stem) they have not been convincingly exhibited in cortex, and it is hard to model gating circuits that can perform arbitrary tasks.
Dr. Magnasco will show that very simple single-layer recurrent neural networks, if excitation and inhibition are finely balanced, have the ability to: store short-term memory in traveling waves of neural activity, “gate” the passage of such traveling waves, and change timescales or lengthscales of neural activity, without any physical change to the synaptic connections, only by sculpting ongoing activity via suitable choice of inputs.
Marcelo O. Magnasco received his Lic. in physics from the National University of La Plata in 1987 and his Ph.D. in physics from the University of Chicago in 1991 under the advice of Leo P. Kadanoff. He conducted postdoctoral research simultaneously with Mitchell J. Feigenbaum at Rockefeller and at NEC Research in Princeton with Albert Libchaber before joining the faculty as assistant professor in 1993. He became associate professor in 1999 and professor in 2003.
MLS lectures are only open to the RU community and will be taking place in Carson Family Auditorium and virtually via Zoom. Virtual participants are required to log in with their RU Zoom account and use their RU email address and password for authentication. We recommend signing out of VPN prior to logging in to the lecture. Please do not share the link or post on social media.
- Open to
- Campus Only