Skip to content

Tutorial 6: Sparse Autoencoders

Paper: Towards Monosemanticity: Decomposing Language Models With Dictionary Learning by Anthropic (2023)

Difficulty: Advanced | Time: 5-6 hours

Status: Coming Soon

Overview

This tutorial will demonstrate how to train and analyze Sparse Autoencoders (SAEs) to decompose polysemantic neurons into interpretable features.

Planned Content

  1. The superposition hypothesis and polysemanticity
  2. Sparse autoencoder architecture and training objective
  3. Training an SAE on model activations using mlxterp's built-in SAE support
  4. Identifying and interpreting learned features
  5. Feature steering: using SAE features for targeted interventions
  6. Limitations and future directions

Implementation Status

This tutorial is tracked in GitHub Issue #6. Contributions welcome!

mlxterp already has SAE training and analysis built-in: - model.train_sae() for training - get_top_features_for_text() for feature analysis

References

  1. Anthropic (2023). Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.

  2. Cunningham et al. (2023). Sparse Autoencoders Find Highly Interpretable Features in Language Models.