1 h 27 min
Seminar, Hosted by Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University
Closing the Discovery Loop with Low-Dimensional Models and Causal Perturbations
Princeton University
Recording
Abstract
Patterned photostimulation optogenetics can perturb neural circuits with unprecedented precision, yet such perturbations often produce weak or counterintuitive effects because the neural code is complex. This talk presents modeling approaches that exploit low-dimensional structure in neural population activity to design stimulation patterns that are both effective and informative.
The first part shows that recurrent neural networks (RNNs) trained to reproduce neural recordings are not sufficiently constrained by low-dimensional, task-driven activity, which undermines their ability to predict perturbation effects reliably. Low-dimensional neural response data instead constrain a latent low-rank connectivity in RNNs, allowing accurate prediction of perturbation effects within the task-relevant subspace. The second part combines optogenetic activation of glomerular patterns in the mouse olfactory bulb with low-dimensional models of glomerular odor responses to relate glomerular activity to olfactory percepts. The olfactory code turns out to be low-dimensional and inherently structured for efficient odor representation, which makes it possible to design novel activation patterns with desired perceptual properties.
Topics
- recurrent neural networks
- Patterned photostimulation, broader topic: photostimulation
- Low-Dimensional Models, search World Wide
- causal perturbations
- neural population activity
- recurrent neural networks (RNNs)
- latent low-rank connectivity
- glomerular odor responses, search World Wide
- mouse olfactory bulb, broader topic: olfactory bulb
- olfactory code, broader topic: code