SAE Training

Reconstruction

Loss

Sometimes, decoder weight sparsity is also included in the loss to include decoder part of sparsity.

Hyperparameters

expansion_factor

Sometimes subtracting bias of decoder to center the input activation to learn meaningful features

SAE Training Techniques

SAE Weight Regularizer

SAE weight initialization

SAE Feature Direction Loss

SAE Training Factors

SAE Distributed Training

Training SAE with L2 loss with several hyperparameters

Circuits Updates - April 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/april-update/index.html#training-saes

Techniques

My best guess at the important tricks for training 1L SAEs — LessWrong

TL;DR: this quickly-written post gives a list of my guesses of the most important parts of training a Sparse Autoencoder on a 1L Transformer, with op…

https://www.lesswrong.com/posts/fifPCos6ddsmJYahD/my-best-guess-at-the-important-tricks-for-training-1l-saes

Specifically, we use a layer 5/6 of the way into the network for GPT-4 series models, and we use layer 8 ( 3/4 of the way) for GPT-2 small. We use a context length of 64 tokens for all experiments.

cdn.openai.com

https://cdn.openai.com/papers/sparse-autoencoders.pdf

SAE Training

Reconstruction

Loss

Hyperparameters

Training SAE with L2 loss with several hyperparameters

Recommendations