Deep neural networks (DNNs) do not yet match the robustness of the human visual system. We hypothesize that the recurrent circuits of the visual system play a key role in the ability of humans to perform visual reasoning. Mental rotation is the task consisting in comparing two 3D objects seen from different viewpoints. The aim of this project is to study mental rotation, as one particular case of visual reasoning. The main objectives of the project are (1) to model the computational mechanisms implemented in the human brain to perform mental rotation (e.g., using psychophysics, neuroimaging, and mathematical modelling), and (2) to improve the robustness of machine learning algorithms by implementing these same principles (e.g., in state-of-the-art pretrained DNNs). We believe that this work could lead to breakthroughs in our understanding of how the human visual system reasons about scenes, as well as significant improvements to the robustness of machine learning algorithms.

2023

2027