Convolutional architectures are cortex-aligned de novo | JoVE Visualize

Area of Science:

Computational Neuroscience
Computer Vision
Artificial Intelligence

Background:

The emergence of cortex-aligned representations in deep neural networks (DNNs) for vision is not fully understood.
Previous research highlighted architectural constraints, but the success of diverse architectures post-pre-training questioned their primary importance.
This study investigates the role of architectural inductive biases in DNNs, particularly convolutional architectures, for visual representation development.

Purpose of the Study:

To determine the underlying factors driving the development of cortex-aligned representations in DNNs.
To assess the significance of architectural inductive biases versus pre-training in achieving visual representation alignment.
To identify specific architectural manipulations that promote cortex-aligned representations.

Main Methods:

Examined DNNs with varied architectures but no pre-training.
Quantified the ability of these networks to predict image representations in primate and human visual cortices.
Investigated the impact of spatial domain compression (pooling) and feature domain expansion (channel increase) on representation alignment.

Main Results:

Convolutional architectures with pooling and feature expansion successfully generated cortex-aligned representations without pre-training.
These dimensionality manipulations were less effective in non-convolutional architectures and in convolutional models with specific architectural modifications.
Architectural inductive biases inherent to convolutional networks were critical for performance gains from feature expansion.

Conclusions:

The inductive biases of convolutional neural networks (CNNs) play a prominent role in developing cortex-aligned visual representations.
These architectural constraints are sufficiently aligned with biological vision principles to allow representation emergence prior to synaptic tuning via experience.
Findings suggest CNN architecture is a significant factor in mimicking biological visual processing, independent of extensive training data.