Sparse Autoencoders Find Highly Interpretable Features in Language Models Article Swipe

MAKE IMPACT

Hoagy Cunningham , Aidan Ewart , Logan Riggs , Robert P. Huben , Lee Sharkey · YOU? · · 2023 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2309.08600

One of the roadblocks to a better understanding of neural networks' internals is \textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task \citep{wang2022interpretability} to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.

Type

preprint

Language

en

Landing Page

http://arxiv.org/abs/2309.08600

PDF

https://arxiv.org/pdf/2309.08600

OA Status

green

Cited By

34

Related Works

10

OpenAlex ID

https://openalex.org/W4386839891

Raw OpenAlex JSON

OpenAlex ID

https://openalex.org/W4386839891

Canonical identifier for this work in OpenAlex

DOI

https://doi.org/10.48550/arxiv.2309.08600

Digital Object Identifier

Title

Sparse Autoencoders Find Highly Interpretable Features in Language Models

Work title

Type

preprint

OpenAlex work type

Language

en

Primary language

Publication year

2023

Year of publication

Publication date

2023-09-15

Full publication date if available

Authors

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert P. Huben, Lee Sharkey

List of authors in order

Landing page

https://arxiv.org/abs/2309.08600

Publisher landing page

PDF URL

https://arxiv.org/pdf/2309.08600

Direct link to full text PDF

Open access

Yes

Whether a free full text is available

OA status

green

Open access status per OpenAlex

OA URL

https://arxiv.org/pdf/2309.08600

Direct OA link when available

Concepts

Interpretability, Computer science, Artificial intelligence, Set (abstract data type), Artificial neural network, Counterfactual thinking, Identification (biology), Machine learning, Deep neural networks, Language model, Superposition principle, Scalability, Task (project management), Natural language processing, Mathematics, Management, Programming language, Epistemology, Biology, Mathematical analysis, Economics, Philosophy, Botany, Database

Top concepts (fields/topics) attached by OpenAlex

Cited by

34

Total citation count in OpenAlex

Citations by year (recent)

2025: 28, 2024: 5, 2023: 1

Per-year citation counts (last 5 years)

Related works (count)

10

Other works algorithmically related by OpenAlex

Full payload