Ph.D. Dissertation Defense: Shoutik Mukherjee

Wednesday, November 29, 2023
3:30 p.m.
AVW 1146
Maria Hoo
301 405 3681
mch@umd.edu

ANNOUNCEMENT: Ph.D. Dissertation Defense 

 
Name: Shoutik Mukherjee
 
Committee:
Prof. Behtash Babadi, Chair/Co-advisor
Prof. Shihab Shamma, Co-advisor
Prof. Jonathan Z. Simon
Prof. Patrick O. Kanold
Prof. Nikolas A. Francis
Prof. Wolfgang Losert, Dean's Representative
 
Date and Time: Wednesday, November 29 2023 at 3:30pm
 
Location: AVW 1146
 
Title: Statistical Models of Neural Computations and Network Interactions in High-Dimensional Neural Data
 
Abstract:
Recent advances in neural recording technologies, like high-density electrodes and two-photon calcium imaging, now enable the simultaneous acquisition of several hundred neurons over large patches of cortex. The availability of high volumes of simultaneously acquired neural activity presents exciting opportunities to study the network-level properties that support the neural code. This dissertation consists of two themes in analyzing network-level neural coding in large populations, particularly in the context of audition. Namely, we address modeling the instantaneous and directed interactions in large neuronal assemblies; and modeling neural computations in the mammalian auditory system.

In the first part of this dissertation, an algorithm for adaptively modeling higher-order coordinated spiking as a discretized mark point process is proposed. Analyzing coordinated spiking involves a large number of possible simultaneous spiking events and covariates. We propose the adaptive Orthogonal Matching Pursuit (AdOMP) to tractably model dynamic higher-order coordination of ensemble spiking. Moreover, we generalize an elegant procedure for constructing confidence intervals for sparsity-regularized estimates to greedy algorithms and subsequently derive an inference framework for detecting facilitation or suppression of coordinated spiking. Application to simulated and experimentally recorded multi-electrode data recordings reveals significant gains over several existing benchmarks.

The second part pertains to functional network analysis of large neuronal ensembles using OMP to impose sparsity constraints on models of neuronal responses. The efficacy of functional network analysis based on greedy model estimation is first demonstrated in two sets of two-photon calcium imaging data of mouse primary auditory cortex. The first dataset was collected during a tone discrimination task, where we additionally show that properties of the functional network structure encode information relevant to the animal's task performance. The second dataset was collected from a cohort of young and aging mice during passive presentations of pure-tones in noise to study aging-related network changes in A1. The constituency of neurons engaged in functional networks changed by age; we characterized these changes and their correspondence to differences in functional network structure. We next demonstrated the efficacy of greedy estimation in functional network analysis in application to electrophysiological spiking recordings across multiple areas of songbird auditory cortex, and present initial findings on interareal network structure differences between responses to tutor songs and non-tutor songs that suggest the learning-related effects on functional networks.

The third part of this dissertation concerns neural system identification. Neurons in ferret primary auditory cortex are known to exhibit stereotypical spectrotemporal specificity in their responses. However, spectrotemporal receptive fields (STRF) measured in non-primary areas can be intricate, reflecting mixed spectrotemporal selectivity, and hence be challenging to interpret. We propose a point process model of spiking responses of neurons in PEG, a secondary auditory area, where neurons' spiking rates are modulated by a high-dimensional biologically inspired stimulus representation. The proposed method is shown to accurately model a neuron's response to speech and artificial stimuli, and offers the interpretation of complex STRFs as the sparse combination of higher-dimensional features. Moreover, comparative analyses between PEG and A1 neurons suggest the role of such an hierarchical model is to facilitate encoding natural stimuli.

The fourth part of this dissertation is a study in computational auditory scene analysis that seeks to model the role of selective attention in binaural segregation within the framework of a temporal coherence model of auditory streaming. Masks can be obtained by clustering cortical features according to their instantaneous coincidences with pitch and interaural cues. We model selective attention by restricting the ranges of pitch or interaural timing differences used to obtain masks, and evaluate the robustness of the selective attention model in comparison to the baseline model that uses all perceptual cues. Selective attention was as robust to noise and reverberation as the baseline, suggesting the proposed attentive temporal coherence model, in the context of prior experimental findings, may describe the computations by which downstream unattended-speaker representations are suppressed in scene analysis.

Finally, the fifth part of this dissertation discusses future directions in studying network interactions in large neural datasets, especially in consideration of current trends towards the adoption of optogenetic stimulation to study neural coding. As a first step in these new directions, a simulation study introducing a reinforcement learning-guided approach to optogenetic stimulation target selection is presented.
 

Audience: Graduate  Faculty 

remind we with google calendar

 

June 2024

SU MO TU WE TH FR SA
26 27 28 29 30 31 1
2 3 4 5 6 7 8
9 10 11 12 13 14 15
16 17 18 19 20 21 22
23 24 25 26 27 28 29
30 1 2 3 4 5 6
Submit an Event