Using Generated Object Reconstructions to Study Object-based Attention.

Object-based attention operates by selecting complete object representations as a fundamental unit, but existing theories tend to focus on bottom-up cues, such as Gestalt principles, and have given relatively little attention to the influence of top-down signals on this process, as well as how bottom-up and top-down factors may interact. Here we propose that Object Reconstruction-guided Attention (ORA) may provide a useful framework to study the interplay between bottom-up and top-down factors in object-based attention. The ORA model encodes object-based representations to reconstruct object location and appearance, and utilizes this reconstructed information to further bias the bottom-up signal in the feedforward pass. The objective of the model is to best explain the input by selecting and reconstructing target objects with the lowest reconstruction error, creating an object-selection bias. Our results demonstrate that this simple reconstruction-based selection principle can support various visual tasks, providing new insights into the brain mechanisms underlying robust object-based attention and visual perception. Future work will extend this work to more naturalistic images and examine the model's correspondence with human behavior.