Gated SAE

Creator

Creator

Seonglae Cho

Created

Created

2024 Oct 24 0:13

Editor

Editor

Seonglae Cho

Edited

Edited

2026 Oct 20 15:56

Refs

Refs

Gated SAE

While L1 loss enforces sparsity, it can excessively reduce important features through shrinkage, making it difficult to properly represent the data.

L1 penalty is only applied during the process of selecting (gating) which features to activate

The degree of activation for selected neurons is determined without L1 penalty

In other words, feature activation remains sparse while preventing the L1 penalty from reducing the actual magnitude of values

Gated SAEs

https://www.lesswrong.com/posts/EWhA4pyfrbdSkCd4G/evaluating-sparse-autoencoders-with-board-game-models

anthropic analysis

Circuits Updates - June 2024

We report a number of developing ideas on the Anthropic interpretability team, which might be of interest to researchers working actively in this space. Some of these are emerging strands of research where we expect to publish more on in the coming months. Others are minor points we wish to share, since we're unlikely to ever write a paper about them.

https://transformer-circuits.pub/2024/june-update/index.html#topk-gated-comparison

Gated SAE

https://arxiv.org/pdf/2404.16014

https://openreview.net/pdf?id=zLBlin2zvW

openai comparison against topk

https://arxiv.org/pdf/2406.04093

Active Budget Can Kill Sensitivity

TopK SAE에서 활성화 feature 수인 \(k\)를 늘리면 reconstruction은 개선되지만, 희귀 feature의 의미적 안정성은 오히려 떨어진다.

https://arxiv.org/pdf/2609.37857

Backlinks

Sparse Autoencoder SAEBench

Recommendations

////////////