Search Papers | Poster Sessions | All Posters
Poster C3 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Robustness to object rotation in humans and deep neural networks
Haider Al-Tahan1,2 (), Farzad Shayanfar, Ehsan Tousi1, Marieke Mur1; 1Western University, 2FAIR, Meta AI
Invariant object recognition, a cornerstone of human vision, enables recognizing objects despite variations in rotations, positions, and scales. To emulate human-like generalization across object transformations, computational models must perform well in this aspect. Deep neural networks (DNNs) are popular computational models for human ventral visual stream processing, though their alignment with human performance on visual tasks remains debated. We examine robustness to object rotation in human adults and pretrained feedforward DNNs. We find that object recognition performance is better preserved in humans than in DNNs, although they show a similar pattern of how performance drops as a function of rotational angle. Furthermore, humans and models make different errors, which suggests different processing strategies. Finally, model architecture minimally influences DNN performance, while DNNs trained on richer visual diets and semi-supervised learning goals excel. Our study suggests that visual diet and learning goals may play an important role in the development of invariant object recognition in humans.
Keywords: Deep neural networks Invariant Object Recognition Ventral Visual Pathway