Deep Learning Models for Visual and Interoceptive Neural Processing

Abstract

Humans perceive their external environment and internal body through complex neural processing in the brain. Exteroception, primarily through vision, enables the brain to understand the world, make decisions, and take actions. Interoception allows the brain to meet physiological needs by sensing and regulating organs, such as the heart, lungs, and gut. Emerging evidence suggests that exteroception and interoception, though seemingly distinct, may share similar mechanisms and architectures for neural computation across complex and hierarchical systems. However, a unified computational understanding remains elusive. My dissertation research leverages advances in deep learning to model both vision (a primary form of exteroception) and interoception. These models are designed based on neuroscience knowledge and are evaluated by comparing their outputs with human behaviors and their internal processes with human brain activity observed with functional magnetic resonance imaging.

My vision model includes two separate streams inspired by the human eyes and visual cortex. One stream processes coarse input from a wide view to control eye movement, while the other stream processes fine input from a narrow view to enable detailed perception. Together, these streams interact to enable scene understanding and object recognition. Results show that this model is more robust against adversarial attack, can generate human-like eye movements, and explains the functional segregation of the dorsal and ventral streams in the human visual cortex.

Similarly, my interoception model includes separate streams to process input from various organs, such as the heart and lungs, inspired by peripheral neural pathways that innervate each organ. These streams converge for integrative processing, forming a holistic representation of the bodily state and predicting its future changes. Results suggest that this model can explain human brain-body interactions and map cortical and subcortical regions involved in interoception.

In summary, my research delivers computational models for understanding biological processes for vision and interoception. These models enable advanced analysis of human neuroimaging data for brain research and potentially for diagnostics of mental illnesses. Beyond the significance in neuroscience, my research also highlights the potential of using neuroscience to inspire artificial intelligence.