An Interpretability Illusion for BERT
read the original abstract
We describe an "interpretability illusion" that arises when analyzing the BERT model. Activations of individual neurons in the network may spuriously appear to encode a single, simple concept, when in fact they are encoding something far more complex. The same effect holds for linear combinations of activations. We trace the source of this illusion to geometric properties of BERT's embedding space as well as the fact that common text corpora represent only narrow slices of possible English sentences. We provide a taxonomy of model-learned concepts and discuss methodological implications for interpretability research, especially the importance of testing hypotheses on multiple data sets.
This paper has not been read by Pith yet.
Forward citations
Cited by 6 Pith papers
-
Localizing Model Behavior with Path Patching
Path patching provides a method to express and quantitatively test hypotheses that neural network behaviors are localized to sets of paths.
-
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
-
Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery
Structurally distinct circuits for literal sequence copying across token frequency bands implement the same computation, shown by broad transfer of band-specific edges, a shared core recovering 99% performance, and in...
-
Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations
SA-GSAE with Bi-Jump-ReLU enables one latent to encode both polarities of anticorrelated features, Pareto-dominating or matching full-width gated SAEs while reducing dead latents by up to 500x on some LLM hookpoints.
-
Scaling and evaluating sparse autoencoders
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
-
Improving Dictionary Learning with Gated Sparse Autoencoders
Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for ...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.