Mapping and editing learned functional geometry inside a CNN (with controls) (github.com)

A new experimental research initiative has unveiled insights into the functional geometry that emerges within convolutional neural networks (CNNs) during training. The project documents the mapping and editing of learned geometric structures in activation space, revealing significant implications for model compression, interpretability, and dynamic behavior of neural networks. Key findings highlight that neurons cluster into structured manifolds, with nearby activations exhibiting functional redundancy. This structured geometry not only aids in safe model consolidation but also suggests that targeted interventions can validate and exploit this geometry for better performance. The research's significance lies in its exploration of the interplay between geometric structure and network behavior, indicating that model compression strategies guided by these geometric insights can outperform conventional methods. Proximity in geometric space correlates with functional overlap, while distant neuron merges could lead to performance degradation. With the completion of various experimental phases—including manifold discovery and causal validation—this project sets the stage for future investigations into plasticity dynamics and cross-architecture validation. The work is shared openly through Google Colab, providing valuable resources for the AI/ML community to build upon.