Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Unlearnability of Symmetry-based Visual Relations by Deep Neural Networks

Guillermo Puebla¹, Jorge Diaz-Ramirez², Domingo Araya³, Gonzalo Fuentes², Jeffrey Bowers⁴; ¹Universidad de Tarapacá, ²Pontificia Universidad Catolica de Chile, ³Pontificia Universidad Católica de Chile, ⁴University of Bristol

Presenter: Guillermo Puebla

Relational reasoning, a fundamental ability of the human mind, remains an important challenge for visual deep neural network (DNN) systems. In this work, we present a new set of visual relational reasoning tasks with varying levels of relational complexity and show that, for a series of convolutional neural networks and vision transformers, a third-order relational task is not learnable---in the sense that extensive hyperparameter search and model training do not yield in-distribution above-chance performance on the test set. Our results highlight the limitations of current visual DNN architectures and suggest a path towards developing reliable methods to discriminate between human and visual DNN system responses.

Topic Area: Computational Models of Vision & Visual Cortex