Speaker: Joshua Gold ARNI WG Multi-resource-cost optimization of neural network models
Title: Flexible information bottlenecks in the brain
Abstract:
The brain must make sense of the world from a constant bombardment of incoming information but using limited computational and other resources. These limitations introduce information bottlenecks in the brain, where the information that is processed is a compressed (capacity-limited) version of the information that is received. I will describe ongoing projects in my lab that aim to better understand the nature of these bottlenecks, in particular their flexibility across conditions and individuals. The first part of the talk will focus on recent experimental findings showing that sensory adaptation, a form of bottleneck evident in the activity of single neurons throughout the brain, can adjust flexibly to recent stimulus statistics to support effective decision-making in dynamic environments. The second part of the talk will focus on recent theoretical work, supported by human psychophysics, that aims to account for this kind of flexibility in terms of balancing inherent trade-offs between the information capacity and accuracy of processes that rely on making inferences about the world from noisy observations. This work also provides new perspectives on optimal behavior, which can be thought of in terms the efficient and effective use of capacity-limited information.
Zoom: Upon request @ [email protected]