monosemanticity
In plain wordsThe property of an internal unit of an AI model that stands for just one understandable thing, which makes it far easier to interpret than a unit that mixes several.
The property of a unit or featurefeature (interpretability), n.: A concept or property that a neural network represents internally — such as 'is French', 'mentions a bridge' or 'this code has a bug' — usually corresponding to a direction in its internal activity. that responds to a single human-interpretable concept across the inputs on which it is active; the opposite of polysemanticitypolysemanticity, n.: When a single artificial neuron responds to several unrelated things — say, bird beaks, bicycle wheels and a particular typeface — so that no single label describes what it does.. Because neurons in language models are rarely monosemantic, research has sought monosemantic units either by redesigning models, as with Anthropic's softmax linear unit (2022), or by decomposing activations with dictionary learningdictionary learning, n.: A method that learns a large vocabulary of basic patterns so that any piece of data can be described using just a few of them. In AI interpretability it builds a dictionary of a model's own concepts.. Anthropic's 'Towards Monosemanticity' (October 2023) trained a sparse autoencodersparse autoencoder, n.: A small add-on network trained to break an AI model's tangled internal signals into a much larger set of simpler pieces, most of which can be labeled with an understandable concept. on a one-layer transformer and found thousands of features — DNA sequences, Arabic script, base64 strings, legal language — far more specific than the model's neurons. 'Scaling Monosemanticity' (May 2024) applied the method to Claude 3 Sonnet, extracting up to about 34 million features, including multilingual, multimodal and safety-relevant ones. Monosemanticity is a matter of degree, judged by human or automated interpretabilityautomated interpretability, n.: Using AI models to help explain other AI models — for instance, having a language model read examples that excite an internal unit and write a description of what it detects. scoring, which can overstate it when only a feature's strongest activations are examined.
Most of the dictionary's features looked monosemantic, but the rarest ones resisted labeling.
Origin
The antonym of 'polysemantic' in its interpretability sense, both words being borrowed from linguistics. In use for neurons by 2022, as in Anthropic's 'Toy Models of Superposition' (September 2022) and Adam Jermyn, Nicholas Schiefer and Evan Hubinger's 'Engineering Monosemanticity in Toy Models' (November 2022). Made prominent by Trenton Bricken, Adly Templeton, Joshua Batson, Chris Olah and colleagues in 'Towards Monosemanticity: Decomposing Language Models With Dictionary Learning' (Transformer Circuits Thread, October 2023), followed by Templeton and colleagues' 'Scaling Monosemanticity' (May 2024).
Related
- sparse autoencodersparse autoencoder, n.: A small add-on network trained to break an AI model's tangled internal signals into a much larger set of simpler pieces, most of which can be labeled with an understandable concept.
- dictionary learningdictionary learning, n.: A method that learns a large vocabulary of basic patterns so that any piece of data can be described using just a few of them. In AI interpretability it builds a dictionary of a model's own concepts.
- Golden Gate ClaudeGolden Gate Claude, prop. n.: A one-day public demo in 2024 in which Anthropic turned up a single internal concept inside its Claude chatbot — the Golden Gate Bridge — so it steered almost every conversation back to the bridge.
- superpositionsuperposition, n.: The way a neural network squeezes more concepts into its internal activity than it has neurons, by letting them overlap — a main reason its insides are hard to read.
- automated interpretabilityautomated interpretability, n.: Using AI models to help explain other AI models — for instance, having a language model read examples that excite an internal unit and write a description of what it detects.
- feature (interpretability)feature (interpretability), n.: A concept or property that a neural network represents internally — such as 'is French', 'mentions a bridge' or 'this code has a bug' — usually corresponding to a direction in its internal activity.
- polysemanticitypolysemanticity, n.: When a single artificial neuron responds to several unrelated things — say, bird beaks, bicycle wheels and a particular typeface — so that no single label describes what it does.
Sources
- Bricken, T., Templeton, A., Batson, J. et al. (2023). Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Transformer Circuits Thread.
- Templeton, A., Conerly, T. et al. (2024). Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread.
Mentioned in
mechanistic interpretabilitymechanistic interpretability, n.: A branch of AI research that tries to reverse-engineer a neural network — working out, piece by piece, which internal parts represent which concepts and how they combine to produce behavior., Golden Gate ClaudeGolden Gate Claude, prop. n.: A one-day public demo in 2024 in which Anthropic turned up a single internal concept inside its Claude chatbot — the Golden Gate Bridge — so it steered almost every conversation back to the bridge., polysemanticitypolysemanticity, n.: When a single artificial neuron responds to several unrelated things — say, bird beaks, bicycle wheels and a particular typeface — so that no single label describes what it does., superpositionsuperposition, n.: The way a neural network squeezes more concepts into its internal activity than it has neurons, by letting them overlap — a main reason its insides are hard to read., automated interpretabilityautomated interpretability, n.: Using AI models to help explain other AI models — for instance, having a language model read examples that excite an internal unit and write a description of what it detects., dictionary learningdictionary learning, n.: A method that learns a large vocabulary of basic patterns so that any piece of data can be described using just a few of them. In AI interpretability it builds a dictionary of a model's own concepts., disentangled representationdisentangled representation, n.: An internal representation in which each part tracks one separate, meaningful factor of the data, such as one number for a face's angle, another for its lighting and another for hair color., feature splittingfeature splitting, n.: What happens when a concept-finding tool is given more room: a broad concept it found before breaks into several narrower ones, much as 'animal' might split into 'dog', 'cat' and 'bird'., Transformer Circuits ThreadTransformer Circuits Thread, prop. n.: Anthropic's online publication for research on how language models work inside, presented as long interactive web articles rather than conventional papers.