"Computational Study of the Mechanism of Attention in Cortical Informat" by Xiang Wang

Date of Award

Fall 2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Interdepartmental Neuroscience Program

First Advisor

Jadi, Monika

Abstract

The animals’ visual perception of natural scenes is influenced by both the external condi-tions of the environment and the internal states. External conditions include the contrast of the stimulus, the crowding of distractors and the complexity of visual features. Internal states include the level of arousal, the deployment of attention, and the goal of behavior. An important and extensively studied internal state is selective attention, which is thought to be critical in solving visual perception problems by increasing the saliency of the attended ob-jects against the background. Electrophysiological studies have indicated attentional modu-lation of single neurons along the hierarchy of visual processing system. However, how attention selection interacts with the stimulus conditions such as contrast and clutter is still debating. More importantly, how different attentional strategies contribute to the behavioral performance remains elusive. Regarding the interaction between visual attention and stimulus contrast, previous literature supported two different patterns: a uniform response gain effect or a biased contrast gain effect. However, both effects were observed by averaging neuronal responses across the whole population of recordings, negating the fact that the sensory cortical region of primate brain is a heterogenous organization with multiple cell classes and distinct laminar com-partments. Therefore, we investigated the attentional modulation pattern of neuronal con-trast responses in a cell-class and layer-specific manner in Chapter 2. We characterized cell classes based on the trough-to-peak duration of the extracellular action potential (EAP) waveform recorded from macaque visual area V4 with k-means clustering algorithm. We found that the attentional effect on contrast responses was not uniform across cell classes or laminar locations, thereby explaining the conflicting results reported in previous studies. Interestingly, the organization of modulation patterns aligned with the demands of the visual processing hierarchy. Finally, we developed a neural circuit model based on the normaliza-tion model of attention that suggested the spatial extent of inhibition as the potential mecha-nism behind the laminar difference in attentional modulation. Our conclusion in Chapter 2 was built on characterization of cell classes given extracellular recordings, which has been a critical technique for studying the heterogeneity of the primate nervous system because in vivo studies in primate do not generally have access to transgen-ic tools for defining cell types based on biomarker expressions. Previous literature cluster-ing EAP waveforms either defined user-specified features or applied network clustering techniques to the raw waveform data. However, a principled way of selecting user-specified features is lacking and clustering on the raw waveforms suffers from the “concen-tration of measure” problem from the high dimensionality. Moreover, the way of selecting optimal number of clusters in different studies were not consistent, making comparisons between clustering studies difficult. To address these issues, we developed a novel cluster-ing technique (VAE-MAP) that combined unsupervised network clustering and extraction of latent representations by variational autoencoder in Chapter 3. We also proposed a prin-cipled way of choosing the optimal cluster number based on soft-voting of multiple cluster-ing quality measures. We demonstrated that our approach not only requires no feature spec-ification but also alleviates the curse of dimensionality and produces more robust clusters compared with previous efforts. Using our clustering pipeline, we showed that both excita-tory and inhibitory neurons from either mice or monkeys correspond to a broad range of EAP waveform shapes, which contradicts the previous view that narrow- and broad-spiking cells are putative inhibitory and excitatory neurons, respectively. To investigate how different attentional modulations affect the behavioral outcome, in Chap-ter 4, we implemented spatial attention in a deep convolutional neural network model of the ventral visual hierarchy and measured its impact on categorization performance on cluttered images with varying contrast levels. We found that the effectiveness of attention depends on the type (enhancement or suppression) and the hierarchical location of neural modulation. Our results suggested that the role of neural enhancement by attention is underestimated without considering the suppression or the hierarchical structure of the visual pathway. Collectively, findings in this thesis deepened our knowledge about the neural mechanisms behind attention-mediated performance improvement. Our results highlighted the necessity of considering cell types, laminar organization, modes of modulation and hierarchy of the primate visual pathway for a more comprehensive understanding of the computational role of attention in information processing.

Download

Share

COinS