Posted Jan 17 2019
Srinivas Turaga, HHMI Janelia Research Campus
"Causally predictive, interpretable, data-driven models of neural computation"
I will describe methods to build data-driven mechanistic neural network models to understand how single neurons and networks of such neurons compute. Can we use calcium imaging and cellular resolution optogenetics perform in vivo measurements of the input-output functions of single neurons, and their connectivity? We have developed a new Bayesian inference method based on variational auto encoders to enable in vivo measurements of synaptic inputs and somatic outputs from dendritic calcium imaging. We also use the same Bayesian inference machinery to infer causal synaptic connectivity in a neural circuit using data from in vivo all-optical circuit interrogation experiments. Electron microscopic reconstructions of neural circuits can provide strong constraints on computation in a neural network. Are these constraints sufficient to predict the tuning properties of individual neurons in the circuit? We constructed a simplified model of the first two stages of the fruit fly visual system, the lamina and medulla. The result is a deep hexagonal lattice convolutional neural network which discovered well-known orientation and direction selectivity properties in T4 neurons and their inputs.
Thursday, January 31, 2019 - 10:00am to Friday, February 1, 2019 - 9:45am
A32 PNI Lecture Hall