Seoyoung Ahn
August 16(Wed) - August 16(Wed), 2023
12pm
86314 & ZOOM
CNIR Seminar
Date: 12pm, Wednesday, August 16th
https://us02web.zoom.us/j/7281426028
회의 ID: 728 142 6028 (password: cnir)
Speaker: Seoyoung Ahn (Stony Brook University)
Title: Generating object-based attention through reconstruction
Abstract: Human can dream and imagine things. These quintessentially human capabilities require the brain to generate percepts. I propose that our generation capabilities likely evolved, not so that we can have more vivid dreams, but rather because generated representations can be used as inference tools by downstream perceptual and cognitive processes. Through a combination of computational modeling and behavioral experiments, I will demonstrate that the generative visual system¾that works by actively reconstructing the most probable object from incomplete retinal input¾can use this reconstruction as top-down attention feedback, thereby supporting cases where object recognition is challenging. The proposed model also better explained human behavioral reaction times and error patterns in these tasks than a comparable feedforward Convolutional Neural Network. I will conclude that reconstruction-guided attention feedback is a plausible mechanism for generating hypotheses about the objects in the world, and that it is key to endowing humans with a robust ability to perceive and recognize objects, bringing a new perspective to object-based attention in this new era of AI.