What problem does it solve? Understanding what individual neurons inside deep vision networks respond to normally requires manual inspection of thousands of activation maps. This Skill automates neuron interpretation by matching neuron activation patterns against natural-language concept sets using CLIP embeddings, producing human-readable descriptions for every unit in a layer. ## Core Features & Use Cases - Automatic Neuron Labeling: Computes cosine similarity between activation-weighted CLIP image embeddings and concept text embeddings to assign each neuron its best-matching concept. - Multi-Layer and Model-Agnostic Dissection: Works with ResNet-50, ResNet-18 (Places-365), Vision Transformers, and custom PyTorch models across multiple layers in one run. - Flexible Probing Setup: Supports Broden and ImageNet validation probing datasets plus bundled 3k, 10k, and 20k English concept sets, with custom datasets and concept files supported. - Use Case: A researcher dissecting ResNet-50 layer4 runs the pipeline with the Broden dataset and 20k concept set, then receives a CSV mapping every neuron to a description like "dog" or "striped" with its similarity score. ## Quick Start Ask the agent to dissect the layers of a pretrained ResNet-50 using the Broden probing dataset and the 20k concept set, saving per-neuron descriptions to a CSV file.
