The project aims to compare human visual attention with CNN attention. School project for Cognitive Science 3 Fall 2018, at KU.
Abstract:
This work explores different attention mechanisms in an object recognition setting with the POET dataset. We introduce a novel approach to constructing heatmaps visualizing human attention, formalize object recognition as a sequential task, and employ an evaluation scheme that proves to distinguish between computational attention mechanisms in relation to human attention. Finally, we use sequential fixations to guide a machine learning model and draw conclusions about the foundational reasons for human effective attention range as a function of eccentricity from the fixation.
Paper : here