Sparse Autoencoder

A Autoencoder trained for:

• Feature learning
• Dimensionality reduction Can be used to extract features from a model. Can be used to remove Polysemanticity

Objective Function

$L=||X-\hat{X}||+\lambda \ast \text{Penalty}(s)$

Usage

• A SAE is itself a neural network that disentagles other neural entworks
• A SAE with 8192 neurons can be used to interpret a LLM with just 512 neurons
• Has been used to interpret Claude 3 Sonnet
• Can be used to find unexpected base64 encodings in small models
• Can be used to find hidden goals
• Used to create Golden Gate Claude

Details

• Encodes into a larger encoding
• Decodes into the original layer