Neural dynamics underlying spatiotemporal cognitive integration

点击翻译按钮获取中文摘要

Project Summary Our ability to visually interpret the world around us depends on rapid bottom-up computations that extract relevant information from the sensory inputs, but it also depends on our accumulated core knowledge about the world providing top-down signals based on prior experience. The goal of this proposal is to study the mechanisms by which visual information is integrated spatially and temporally to combine bottom-up and top- down knowledge. Towards this goal, we combine behavioral measurements, invasive neurophysiological recordings, invasive electrical stimulation, and computational models. We focus on the ubiquitous challenge of visual search, exemplified by searching for your phone using exclusively visual cues. The behavioral data will provide critical constraints about human integrative abilities, particularly through eye movements and the dynamics of recognition and object location. The invasive neurophysiological data will provide high spatiotemporal resolution of neural activity along the inferior temporal cortex and the interactions with the pre- frontal cortex, which are hypothesized to be critical for conveying the type of top-down signals required for recognition and attention modulation during visual search. Ultimately, a central goal of our proposal is to formalize our understanding of these integrative processes via a quantitative computational model. This computational model should be able to capture the behavioral and physiological results and provide testable predictions. During the current award, we have made progress towards elucidating the mechanisms underlying pattern completion used by the visual system to infer the identity of objects from partial information, the effects of contextual information during object recognition, and computational models of visual search. We have strong preliminary evidence that suggests that state-of-the-art purely bottom-up theories of recognition instantiated by deep convolutional networks cannot explain human behavior and physiology. Therefore, the proposed work aims to establish a strong computational, behavioral and physiological framework that merges bottom-up and top-down processing. Furthermore, we will move beyond correlative measures by using electrical stimulation to stress test the models and establish causal links between key nodes in the circuitry and visual search behavior. Understanding the neural mechanisms by which core knowledge is incorporated into sensory processing is arguably one of the greatest challenges in Cognitive Science and may have important implications for many neurological and psychiatric conditions that are characterized by dysfunctional top-down signaling and remain poorly understood.