sparse-autoencoder-training

To train sparse autoencoders, you configure sparsity penalties, expansion factors, and warmup schedules to decompose dense neural network activations into sparse, human-interpretable features. This process also requires mitigating dead features and monitoring reconstruction quality.

Mechanistic interpretability uses sparse autoencoders for feature discovery by decomposing dense activations into interpretable components. This allows researchers to analyze superposition, identify monosemantic features, and understand how representations are encoded within language models.

Yes, you can steer model behavior using decoder directions extracted from a pretrained SAE. The workflow supports loading existing sparse autoencoders to ablate specific features, attribute predictions, and actively steer TransformerLens-backed language models.

Yes, custom SAE training requires SAELens and TransformerLens alongside PyTorch. This stack provides the necessary support for encoding activations, decoding reconstructions, managing sparsity regularization, and analyzing feature activations across language models.

The best way to analyze SAE feature activations is to load a pretrained sparse autoencoder for a chosen model layer and summarize the top activating features for your specific prompt. This allows direct inspection of interpretable components within the model.

You need dead-feature mitigation when training sparse autoencoders to prevent features from becoming permanently inactive during sparsity regularization. Managing dead features alongside warmup schedules ensures high reconstruction quality and effective feature discovery.