Module committee 

Query-by-committee — query the points where an ensemble disagrees. Disagreement is the signal: where independently-trained models diverge, a human label resolves the most uncertainty.

Two soft measures (over each member’s predicted distribution) and one hard measure (over each member’s single vote), all per-sample; plus a pool ranking.

Functions

average_kl_disagreement

Average KL disagreement: mean KL(member ‖ consensus) over the committee. This is the canonical soft-QBC measure — it is large precisely when members are confident but about different classes (genuine disagreement), not merely diffuse.

consensus

The committee consensus distribution: the mean of the members’ predicted distributions for one sample. members is n_members × n_classes.

consensus_entropy

Consensus entropy (soft vote entropy): entropy of the consensus distribution. Diffuse consensus ⇒ informative.

rank_by_disagreement

Rank pool indices by average-KL disagreement, most-informative first. pool is n_samples committees, each n_members × n_classes.

vote_entropy

Vote entropy over hard votes: entropy of the distribution of committee votes across n_classes. Maximal when the committee splits evenly. votes are class indices, one per member.