Deep Neural Networks as a Computational Model for Human Shape Sensitivity.

Area of Science:

• Cognitive Science
• Computer Vision
• Neuroscience

Background:

• Object recognition theories emphasize shape but lack consensus on representation.
• Previous computational models of shape perception have struggled with realistic stimuli.
• Recent advances show deep neural networks (DNNs) capture aspects of human object perception.

Purpose of the Study:

• To investigate if deep neural networks (DNNs) develop human-like shape representations.
• To test if DNNs trained for object recognition can explain human shape judgments.
• To explore DNN sensitivity to shape features and non-accidental properties.

Main Methods:

• Training convolutional DNNs on natural photographs for generic object recognition.
• Evaluating DNNs against human behavioral and neural data for shape judgments.
• Testing DNNs on stimulus sets dissociating shape from category membership.

Main Results:

• DNNs trained on object recognition spontaneously develop sensitivity to human-like shape features.
• DNNs successfully explain human shape judgments on benchmark datasets where prior models failed.
• Complex DNN architectures capture human shape sensitivity even when shape and category are dissociated.

Conclusions:

• Convolutional neural networks learn perceptually accurate shape representations, not just physically correct object categories.
• DNNs offer a promising computational framework for understanding human and primate object and shape representation.
• Future models may benefit from training deep architectures on multiple tasks, mirroring human development.