Overview
Siwei Wang is an assistant professor in the Department of Neurobiology and Behavior at Stony Brook University. Her previous work has shown how information processing during survival-critical behaviors sculpts early neural systems. She uses information theory, Bayesian inference, and representation learning methods to investigate how neural representation evolves along the sensory-to-behavior arc to enable effective future action planning based on past experiences.
Title: Search for common principles for efficient representation in both neural coding and AI
Abstract:
In the wild, survival often depends on anticipating what happens next. A fly must escape a predator’s strike before its visual system even finishes processing the threat. A salamander needs to track prey moving through completely different environments as it migrates from aquatic to terrestrial environments. How neural systems extract actionable insights within a split-second while neural processing is inherently laggy? Given this temporal constraint, predictive information is one such common computation. By encoding patterns predictive of future maneuvers, specialized motion sensitive neurons in the fly visual system enables them to steer animals away from visual threat within 40 ms, half the time as our eye blinks. Similarly, in the salamander retina, the encoding of predictive information sculpts the population code to establish a low-dimensional, transferable representation that seamlessly generalizes across radically different natural scenes. In addition, when this representation combines static and dynamic features synergistically to anticipate future states, its compression rivals the most state-of-the-art video compression algorithms.
The above insights from biology inspires new theories in AI. By extracting the general principles for the above biological efficient representations, the newest work from my group demonstrates that such principle also applies to feature engineering in contrastive learning writ large. The overarching goal of this walk is to establish a virtuous cycle where developing biologically plausible AI systems helps interrogate unknown encoding mechanisms in the brain, while these discovered neural principles simultaneously inspire more efficient computational architecture.
The NSF-Simons National Institute for Theory and Mathematics in Biology Seminar Series aims to bring together a mix of mathematicians and biologists to foster discussion and collaboration between the two fields. The seminar series will take place on Fridays from 10am – 11am at the NITMB in the John Hancock Center in downtown Chicago. There will be both an in-person and virtual component.
More information: https://www.nitmb.org/nitmb-seminar-series