Poster Z17 - Identifying Critical Features in Neural Population Data for Sound Source Localization Using Machine Learning Approaches

Introduction: Mammals possess a remarkable ability to accurately locate sound sources, which relies on the integrity of the primary auditory cortex (A1). Previous research has aimed to identify a cortical map that faithfully encodes auditory space at the level of A1, but without much success. This study seeks to bridge the gap between how spatial location is represented within the auditory cortex and the precise behavioral performance of sound localization.

Materials and

Methods: We applied multichannel electrophysiological recording from the auditory cortex when the mouse was listening to sounds coming from different locations.
We applied band-pass filters to get multi-unit activity and local field potential.
We used multiple machine learning algorithms, including unsupervised learning like Principal Component Analysis to visualize data, and supervised learning like K-Nearest Neighbor, Support Vector Machine and 3D Convolutional Neural Network to do classification on multidimensional time series data.

Results, Conclusions, and Discussions: Both spike rate and timing, along with LFP signals in the delta (0-5 Hz) and gamma (40-120 Hz) frequency bands, are critical for encoding sound source locations.

The overall magnitude of the event-related potential in the gamma band is highly predictive of how well that electrode site will be in predicting sound source location.

Both LFPs and spike data contain information about sound source locations, but they are not redundant.

Acknowledgements (Optional):