| Cole Hurwitz |
| Tue 12 Mar 2019, 11:00 - 12:00 |
| IF 4.31/4.33 |
If you have a question about this talk, please contact: Gareth Beedham (gbeedham)
Cole Hurwitz
Point Source Localization in Extracellular Recordings using Amortized Variational Inference
Extracellular recordings using modern, dense probes provide detailed footprints of action potentials (spikes) from thousands of neurons simultaneously. Inferring the activity of single neurons, however, is a complex blind source separation problem, complicated both by the high intrinsic data dimensionality and large data volume. Here we present a new generative model for localizing the source of individual spikes given observed electrical traces. Assuming an exponential decay of signal amplitudes from an unknown point source location of the spike, we perform MCMC sampling-based Bayesian inference on realistic simulated extracellular action potentials to verify the model assumptions and to understand its limitations. In a next step, we perform rigorous data augmentation to improve localization performance and then re-implement the model as a variational autoencoder to improve the inference speed by orders of magnitude. We show that our final model is superior to heuristic methods such as center of mass and can also be applied to large extracellular datasets inaccessible to traditional inference techniques.