|M.Sc Student||Avinu Maya|
|Subject||Analyzing Neuronal Signals using Geometric Methods|
|Department||Department of Electrical Engineering||Supervisors||Professor Ronen Talmon|
|Professor Ron Meir|
|Full Thesis text|
Analysis of neuronal activity recordings constitutes a major step towards understanding the dynamics of cortical networks and decoding the brain activity. Such datasets are typically highly complex and cannot be addressed by traditional analysis and learning methods. In this research, we analyze two-photon recordings from multiple identified neurons along time from numerous trials spanning several days of experiments. We view the data as a three-dimensional tensor, whose dimensions represent neurons, time and trials, and assume that there exists an underlying organization on each dimension. Our goal is to uncover the intrinsic organization in a manner that respects the co-dependencies between the tensor dimensions. Following recent work, we organize the data tensor in coupled hierarchical structures based on informed metrics, which are iteratively built from the data. Specifically, we propose a purely data-driven procedure, in which multiple metrics obtained under different configurations are aggregated into a single consensus robust representation. The gist of the algorithm is based on the concept of alternating diffusion, which extracts the common information and discards nuisance effects. We show that our method is insensitive to parameter variations and data perturbations. We then demonstrate the performance of the algorithm on in-vivo neuronal data. In additional, we show application to unsupervised segmentation of speech signals into voiced, unvoiced and silence frames.