Neural Geometry in Vision Models with Block-Sparse Featurizers
7 hours ago
- #Computer Vision
- #Neural Networks
- #Interpretability
- Block-Sparse Featurizers (BSF) decompose model activations into multidimensional subspaces instead of single directions, capturing concepts as manifolds.
- BSFs outperform standard sparse autoencoders (SAEs) in recovering interpretable features and explaining model activations more efficiently, as shown in toy models and vision models like DINOv3.
- Features found by BSFs enable fine-grained steering, allowing exploration of concept variations, such as different types of trees or pretzels, by moving within the multidimensional subspaces.
- Most concepts in vision models are multidimensional, typically two- to four-dimensional, challenging the assumption that concepts are represented as one-dimensional lines.
- BSFs reveal continuous structures in neural networks, such as curve detectors in InceptionV1, uncovering symmetries like Fourier harmonics that were previously fragmented.
- The principle of block-sparsity is key, emphasizing that subspaces fire together, and different BSF variants (vanilla, Grassmannian, group-lasso) implement this with varying encoder/decoder designs.
- Future work may involve adapting featurizers to different domains, as block-sparsity may not be universally optimal, and interpretability should align tools with the inherent geometry of model representations.