Layerwise complexity-matched learning yields an improved model of cortical area V2.

Area of Science:

• Computational neuroscience
• Computer vision
• Machine learning

Background:

• Human visual pattern recognition involves hierarchical transformations in the ventral visual cortex.
• Deep neural networks excel at late-stage visual processing but struggle with early stages and biological plausibility.
• Gradient backpropagation, common in deep learning, is considered biologically implausible.

Purpose of the Study:

• To develop a biologically plausible, bottom-up self-supervised training methodology for neural networks.
• To improve the modeling of early visual processing stages.
• To enhance generalization and biological alignment in artificial vision models.

Main Methods:

• Introduced layerwise complexity-matched learning (LCL) using self-supervised training on successive layers.
• Maximized feature similarity for locally-deformed image patches and decorrelated features across different images.
• Adjusted deformation amplitudes proportionally to receptive field sizes at each layer to match task complexity.

Main Results:

• The LCL formulation produced a two-stage model (LCL-V2) with improved alignment to primate area V2 selectivity and neural activity.
• Complexity-matched learning was identified as a key factor for enhanced biological alignment.
• The LCL-V2 front-end improved object recognition models (LCL-V2Net) in out-of-distribution generalization and human behavior alignment.

Conclusions:

• Layerwise complexity-matched learning offers a more biologically plausible and effective approach to training artificial vision systems.
• This method bridges the gap between deep learning models and early-stage visual processing in humans.
• The developed models show superior performance and better mimic human visual capabilities.