Modeling Rapid Contextual Learning in the Visual Cortex with Fast-Weight Deep Autoencoder Networks

Area of Science:

• Computational neuroscience
• Deep learning
• Visual processing

Background:

• Early visual cortex learns global context via local recurrent interactions.
• This learning sparsifies neural responses and reduces mean activity for familiar contexts.
• Recurrent neural circuits reshape neural manifolds, improving robustness and invariance.

Purpose of the Study:

• Investigate how familiarity training induces global context sensitivity in early deep neural network layers.
• Explore the role of fast weights in rapid learning using Low-Rank Adaptation (LoRA).
• Model the functional consequences of rapid global context learning in the brain.

Main Methods:

• Employed a Vision Transformer (ViT)-based autoencoder for functional investigation.
• Utilized Low-Rank Adaptation (LoRA) to implement fast weights within Transformer layers.
• Analyzed self-attention mechanisms and latent representations during familiarity training.

Main Results:

• ViT autoencoder's self-attention performed manifold transforms similar to neural models of familiarity.
• Familiarity training aligned early layer representations with global context information in top layers.
• Self-attention broadened its scope to encompass more image details, not just object recognition features.
• LoRA-based fast weights significantly amplified these familiarity-induced effects.

Conclusions:

• Familiarity training can introduce global sensitivity to earlier layers in hierarchical networks.
• A hybrid fast-and-slow weight architecture offers a computational model for rapid global context learning.
• This approach provides insights into the neurobiological mechanisms of visual context processing.