Calvin Lab Rm 116
On Inferring Strongly Recurrent Circuits from Neural Activity
Understanding the mechanisms of neural computation requires knowledge of the underlying circuitry. Because of the difficulty of directly measuring complete circuit wiring diagrams, there has long been interest in estimating them from neural recordings. In this talk, I will show that such even sophisticated inference algorithms applied to large volumes of data for every node in the circuit are biased toward inferring connections between unconnected neurons with high correlations, as can happen in strongly recurrent circuits, representing a failure to fully "explain away". This effect occurs when there is a mismatch between the true network dynamics and the generative model assumed for inference, an inevitable situation when we model the real world. The problem is greatly exacerbated in the typical case of the recorded neurons being a subset of the network, suggesting that connectivity estimates in strongly connected networks should be treated cautiously.
If you would like to give one of the weekly seminars on the Brain & Computation program, please fill out the survey at https://goo.gl/forms/zRlEuwjzdP6MsB6I2