Compact deep neural network models of the visual cortex.

Area of Science:

• Computational neuroscience
• Machine learning
• Primate visual system

Background:

• Deep neural networks (DNNs) are powerful tools for modeling neural responses but are often large and complex.
• Understanding the computations within the primate visual cortex requires predictive models.

Purpose of the Study:

• To develop predictive and parsimonious DNN models of the primate visual cortex.
• To investigate the internal workings of compressed DNN models.

Main Methods:

• Adaptive closed-loop experiments combining data collection and DNN model training.
• Compression of a large DNN model (60 million parameters) to identify compact models (5,000x fewer parameters).
• Analysis of compact models to uncover computational motifs and mechanisms.

Main Results:

• Achieved highly predictive DNN models for macaque visual area V4.
• Successfully compressed a large DNN model, retaining high accuracy with significantly fewer parameters.
• Discovered a computational motif of shared early filters followed by specialized feature selectivity consolidation.
• Identified a circuit hypothesis for dot-selective V4 neurons.
• Demonstrated strong model compression for visual areas V1 and IT, suggesting a general principle.

Conclusions:

• Large DNNs are not always necessary for predicting individual neuron responses.
• A modeling framework balancing prediction and parsimony is established.
• A general computational principle of feature selectivity consolidation exists in the visual cortex.