Compression-enabled interpretability of voxelwise encoding models.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are effective for voxelwise modeling of brain activity from natural movies.
• High parameter counts in CNNs hinder interpretability and stability of these predictive models.

Purpose of the Study:

• To investigate if model compression can enhance the interpretability and stability of CNN-based voxelwise models.
• To assess if compression techniques maintain or improve predictive accuracy.

Main Methods:

• Applied filter/connection pruning and receptive field compression to CNN models.
• Utilized principal component analysis for dimensionality reduction.
• Evaluated model performance on a hold-out test set for visual stimulus identification.

Main Results:

• Model compression improved the accuracy of visual stimulus identification.
• Compressed models provided more stable interpretations of voxelwise pattern selectivity.
• Receptive field compression resulted in larger, more centralized population receptive fields along the ventral visual pathway.

Conclusions:

• Model compression is a viable strategy for developing more interpretable and stable voxelwise CNN models.
• Compression techniques enhance the understanding of neural representations in the visual cortex.
• Optimized receptive fields suggest shifts in visual processing along the ventral stream.