Sparse Crosscoders for Cross Layer Features and Model Diffing
Sparse Crosscoders for Cross Layer Features and Model Diffing
This research update from Anthropic introduces sparse crosscoders, a novel deep learning technique designed to understand and compare complex neural networks. Unlike traditional sparse autoencoders, which focus on individual layers, crosscoders analyze features distributed across multiple layers, revealing hidden connections and simplifying circuit analysis. These crosscoders enable "model diffing," which compares features and circuits between different models, including finetuned variations and models with different architectures. By highlighting shared and model-specific features, crosscoders offer insights into how models learn and evolve, potentially aiding in the development of safer and more interpretable AI systems.
Paper: https://transformer-circuits.pub/2024...
This podcast is generated using NotebookLM for the research purpose.