High-performing neural network models of visual cortex benefit from high latent dimensionality

Area of Science:

• Computational Neuroscience
• Machine Learning
• Computer Vision

Background:

• Geometric descriptions of deep neural networks (DNNs) offer insights into computational models used in neuroscience.
• A prevailing hypothesis suggests optimal DNNs utilize low-dimensional representations for invariance and robustness.
• This implies that superior models of the visual cortex should exhibit lower-dimensional geometries.

Purpose of the Study:

• To investigate the geometry of DNN models of the visual cortex.
• To quantify the latent dimensionality of natural image representations within these models.
• To correlate representational geometry with generalization performance in predicting neural responses.

Main Methods:

• Examined DNN models of the visual cortex.
• Quantified latent dimensionality of natural image representations.
• Assessed generalization performance using monkey electrophysiology and human fMRI data for held-out stimuli.

Main Results:

• Found a trend opposite to the prevailing hypothesis: higher-dimensional image subspaces correlated with better generalization.
• High dimensionality predicted cortical responses more accurately in both monkey and human datasets.
• Higher dimensional representations were associated with improved performance in learning new stimulus categories.

Conclusions:

• High-dimensional geometry confers computational benefits to DNN models of the visual cortex.
• Increased dimensionality enhances generalization capabilities beyond training domains.
• Challenges the notion that low-dimensional compression is universally optimal for visual cortex models.