liveBook · Manning

Sparse autoencoder

A sparse autoencoder is an autoencoder trained to produce representations in which most encoded units are inactive or near zero.

How 2 Manning books explain it

Rearchitecting LLMs

A sparse autoencoder is an autoencoder that reconstructs neural activations through a wider intermediate representation constrained to activate sparsely.

How it is used
Supports mechanistic interpretability by isolating interpretable features for more precise concept-level intervention.
How deep it goes
Defines sparse reconstruction and feature intervention, but does not provide training procedures or implementation details.
Examples
Uses feature isolation in a single-layer Transformer and Claude 3, plus Golden Gate Bridge steering.
What you can do after
Understand how sparse features can make concept-specific inference steering more precise.

Read how Rearchitecting LLMs explains sparse autoencoder →

Data Without Labels

A sparse autoencoder is an autoencoder trained to produce representations in which most encoded units are inactive or near zero.

How it is used
Addresses overfitting while learning sparse representations for encoding, dimensionality reduction, and data compression.
How deep it goes
Defines the activation penalty and its rationale, but does not provide a loss formula or sparse-model implementation.
Examples
Compares autoencoders with PCA through a figure contrasting nonlinear and linear dimensionality reduction.
What you can do after
Evaluate how activation penalties reduce hidden-unit activity and distinguish sparse regularization from bottleneck-based regularization.

Read how Data Without Labels explains sparse autoencoder →

Related concepts

Questions about sparse autoencoder

What is sparse autoencoder?

A sparse autoencoder is an autoencoder trained to produce representations in which most encoded units are inactive or near zero.

Which Manning books explain sparse autoencoder?

Rearchitecting LLMs and Data Without Labels.

What is the difference between sparse autoencoder and logit lens?

A sparse autoencoder is an autoencoder trained to produce representations in which most encoded units are inactive or near zero. The logit lens is an interpretability technique that projects intermediate hidden states through a model's final normalization and prediction head to inspect layer-wise predictions.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.