Gaurav Malhotra

Human shape representations are not an emergent property of learning to classify objects

This study shows that supervised CNNs do not develop human-like sensitivity to relational shape changes when classifying objects.

The role of capacity constraints in Convolutional Neural Networks for learning random versus natural data

This paper investigates how capacity constraints in CNNs affect learning of random versus natural data, finding that internal noise most effectively reduces performance on random data to chance levels.

Feature blindness: a challenge for understanding and modelling visual object recognition

This paper investigates shape-bias in humans and CNNs, finding that humans rely on shape despite non-shape features being more diagnostic, while CNNs adapt to predictive features.

Obstacles to inferring mechanistic similarity usingRepresentational Similarity Analysis

The paper identifies issues with Representational Similarity Analysis (RSA) when comparing neural systems, showing how confounds in stimuli can lead to misleading similarity scores between different systems.

Biological convolutions improve DNN robustness to noise and generalisation

This paper shows that adding biological filter banks improves DNN robustness to noise and generalisation to out-of-domain images.

What do adversarial images tell us about human vision?

This study reanalyzes adversarial image data to show weak and variable agreement between humans and DCNNs, challenging the claim that DCNNs are good models of human vision.