Learning to attend in a brain-inspired deep neural network
Halls department, Hall 3
Wednesday, 26 December 2018
14:00 - 15:00
Visual attention enables prioritization, selection and further processing of visual inputs for the purpose of achieving behavioral goals, but how is this attention control learned? How does it mechanistically serve the selection of correct actions? Recent methods in machine learning have engaged these difficult questions, and showed that including attention as a component results in improved model accuracy and interpretability. These models have yielded better performance in a range of tasks from object classification to caption generation and language translation. However, the attention mechanism in these models were not informed, beyond a broad concept of "attending" (selectively integrating information across time and space), by cognitive and neural findings in primate attention mechanisms. Here we leverage the potential of these perspectives by introducing ATTNet, a deep neural network model of the ATTention Network with layers in this network corresponding to key brain areas in ventral and dorsal visual pathways (processing 'what' and 'where' object is perceived in the brain) involved in prioritizing visual information and planning eye-movements. Using reinforcement learning, ATTNet is trained to detect targets in images. Both subjects and ATTNet showed evidence for attention being preferentially directed to target goals, behaviorally measured as eye-movements' guidance to the targets. More fundamentally, ATTNet learned to spatially route its visual inputs so as to maximize target detection success and reward, and in so doing learned to shift its attention. This work makes a step toward a better understanding of the role of attention in the brain and other computational systems that can lead to AI systems that learn and think more like people.
Hossein Adeli is a research scientist at Stony Brook University in the departments of Psychology and Computer Science. He is interested in understanding the role of attention control and selective routing of information in both Cognitive and Computational systems. His research interests also include developing Machine learning and AI techniques for different applications inspired by Cognitive and Neural Systems. He received his PhD in Cognitive Science from Stony Brook University working in Eye Cognition and Computer Vision Labs. He also has a master's degree in Computer Science and received his bachelors degree in Electrical Engineering form Sharif University of Technology.