Increasing neural network robustness improves match to macaque V1 eigenspectrum, spatial frequency preference and...

Area of Science:

• Computational neuroscience
• Machine learning
• Computer vision

Background:

• Task-optimized convolutional neural networks (CNNs) mimic the brain's ventral visual stream but are brittle to imperceptible image perturbations.
• Understanding this brittleness is crucial for developing more reliable AI systems.

Purpose of the Study:

• Investigate the relationship between model robustness and neural response representations.
• Compare computational models with neurophysiological data from the primary visual cortex (V1).

Main Methods:

• Analyzed eigenspectrum power law exponents of neural responses in robust and non-robust models.
• Examined spatial frequency tuning of artificial neurons in models.
• Compared model representations with neurophysiological data from mouse and macaque V1.

Main Results:

• Neural responses in mouse and macaque V1 exhibit eigenspectra with power law exponents >= 1, aligning with theoretical robustness predictions.
• CNN eigenspectra decay slower than neurophysiological data; robust models show faster decay and higher exponents.
• Robust models demonstrate spatial frequency tuning more aligned with macaque V1, outperforming non-robust models in predicting V1 activity.

Conclusions:

• Slow eigenspectrum decay in models suggests encoding of fine features, contributing to brittleness.
• Aligning model representations with V1's spatial frequency tuning and eigenspectrum properties enhances robustness and predictive power.
• Penalizing slow eigenspectra or biasing models towards lower spatial frequencies may improve AI robustness and V1 predictivity.