Grant: Discovery Launch Supplement

The exceptional ability of the human brain to recognize objects, even in demanding situations, is undeniably remarkable. Take, for example, the capacity to identify a friend's face in a poorly lit room or to notice a partially hidden pedestrian behind a parked car while driving, enabling us to navigate complex environments safely. How does the brain effortlessly bridge gaps in information, ensuring precise interpretation of our surroundings? My long-term vision for my research program is to explore neural feedback connections in the visual cortex, aiming to understand their functional roles in object recognition in challenging conditions and to apply insights to artificial intelligence (AI) models to make them robust to adversarial attacks. Using behavioral, neuroimaging (fMRI, EEG), machine learning, and deep learning techniques, we seek to map the spatiotemporal dynamics of human brain recurrent processes in visual cognition. Working towards this long-term vision, over the next five years, I will focus on the following short-term objectives: Objective 1. To map the spatiotemporal neural dynamics of feedback processes in visual cognition under degraded visual conditions. By conducting neuroimaging experiments (fMRI, EEG) to investigate neural feedback mechanisms during challenging visual conditions like masking and occlusion, we identify the brain regions involved and the duration and sequencing of processes using multivariate pattern analysis techniques. Objective 2. To investigate the functional role of neural feedback connections in visual cognition using computational models. We design computational models and experiments to assess the performance of deep neural network (DNN) models for visual recognition under challenging conditions, similar to our human experiments. This helps us identify conditions where DNNs fail and how local recurrent and top-down feedback connections can enhance their performance. Objective 3. To investigate the role of feedback processes in the absence of feedforward vision. We investigate feedback mechanisms in mental imagery, which lacks feedforward visual input. We determine when and where mental imagery occurs in the brain, its spatiotemporal neural mechanisms, and explore how feedback from memory regions to the visual cortex constructs the visual experience. Further we model this task with deep learning and generative model techniques, and study the computational principles of neural feedback processes in the absence of feedforward visual input. To achieve these goals, we will bridge both high spatial and high temporal resolution neural data (fMRI and EEG) with deep learning-based computational models. We will employ machine learning techniques for neural decodings and use representational similarity analysis to link fMRI with EEG data and computational models. My research program at the AI and visual neuroscience intersection will contribute to our understanding of human brain and advance the AI technology.