FeatLens – One API to visualize features from any vision backbone
FeatLens is a unified API that enables visualization of features from any vision backbone, simplifying the process of inspecting and understanding learned representations across different neural network architectures.
Background
- FeatLens is an open-source tool that lets developers visualize what neural networks "see" inside their layers — the feature maps or intermediate representations — using a single API that works across many different vision models (e.g., CNNs like ResNet, vision transformers like ViT, etc.).
- Feature maps are the internal activations of a neural network; visualizing them helps researchers debug why a model makes certain predictions, detect biases, or understand which image regions the model focuses on.
- Existing tools often require custom code for each model architecture. FeatLens aims to unify that process, making model interpretability more accessible to ML engineers and researchers.