Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
This work addresses the challenge of mechanistic interpretability for researchers by providing a new unsupervised method to decompose representation spaces, though it appears incremental as it builds on existing interpretability concepts.
The paper tackles the problem of understanding how neural models organize different input aspects into separate subspaces within their representation space, and demonstrates that unsupervised neighbor distance minimization can find interpretable, non-basis-aligned subspaces, with quantitative experiments on GPT-2 showing strong connections to known circuits and scalability to 2B models.
Understanding internal representations of neural models is a core interest of mechanistic interpretability. Due to its large dimensionality, the representation space can encode various aspects about inputs. To what extent are different aspects organized and encoded in separate subspaces? Is it possible to find these ``natural'' subspaces in a purely unsupervised way? Somewhat surprisingly, we can indeed achieve this and find interpretable subspaces by a seemingly unrelated training objective. Our method, neighbor distance minimization (NDM), learns non-basis-aligned subspaces in an unsupervised manner. Qualitative analysis shows subspaces are interpretable in many cases, and encoded information in obtained subspaces tends to share the same abstract concept across different inputs, making such subspaces similar to ``variables'' used by the model. We also conduct quantitative experiments using known circuits in GPT-2; results show a strong connection between subspaces and circuit variables. We also provide evidence showing scalability to 2B models by finding separate subspaces mediating context and parametric knowledge routing. Viewed more broadly, our findings offer a new perspective on understanding model internals and building circuits.