Codebook Features: Sparse and Discrete Interpretability for Neural Networks
π Description
CODEBOOK FEATURES : SPARSE AND DISCRETE INTERPRETABILITY FOR NEURAL NETWORKS Alex Tamkin Anthropic?Mohammad Taufeeque FAR AINoah D. Goodman Stanford University ABSTRACT Understanding neural networks is challenging in part because of the dense, con- tinuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features . Code- book features are produced by finetuning neural networks with vector quantization bottlenecks at each layer, producing a network whose hidden features are the sum of a small number of discrete vector codes chosen from a larger codebook. Sur- prisingly, we find that neural networks can operate under this extreme bottleneck with only modest degradation in performance. This sparse, discrete bottleneck also provides an intuitive way of controlling neural network behavior: first, find codes that activate ...
π Game Impacts
Not verified in gameWhich variables this event was proposed to move, and in which direction. The magnitudes are held in the corpus but are not shown here, because they have not been verified against the shipped game. They come from pdoom-data. They describe what an event was proposed to do, not what the shipped game does with it. Most events in the corpus are flavour: they are shown for colour and do not move any game variable. Only a small minority reach the systems below, and several of the variables listed here are not read by the game at all yet. Treat this table as a design proposal under review, not as a measurement of play. Corrections and arguments are welcome — the suggestion links at the foot of this page go straight to the data repo.
| Variable | Direction | Condition |
|---|---|---|
| Research | proposed: up | Always |
| Papers | proposed: up | Always |
| Vibey Doom | proposed: up | Always |
π Reactions
π Sources
π€ Found an Issue?
This event data is sourced from the pdoom-data repository. If you notice errors or want to suggest improvements:
GitHub Issue (Preferred) π§ Email (No GitHub)