Query by committee is an early active learning rule. Train a committee of student models (here perceptrons) on the labelled data so far, then ask for the label of the input on which the committee disagrees most. The paper studies it on two toy problems: the high-low game and a perceptron learning another perceptron.
Findings
- Each query keeps a finite information gain as learning goes on, while for randomly chosen inputs the gain goes to zero.
- As a result generalisation error falls exponentially with the number of examples, against a much slower inverse power law for random inputs.
- The authors suggest a finite asymptotic information gain may be the mark of a good query algorithm.
Limits
- I read a summary of the abstract only, through an index page, and checked the bibliographic record on Crossref.
- Toy problems with realisable targets and noise-free labels.
Relevance
The oldest version of “disagreement is information”: where the committee splits is where a new label is worth most. For a robot the query can be physical: in GoGet the agent treats disagreement as a cue to move closer before answering, which is query by committee with the robot’s body as the query. See active perception and multi-robot multimodal perception.
Links
- DOI
- Subfield: Robot learning