Query-by-Committee (QBC) trains K classifiers on bootstrap resamples of the current labeled set. Each unlabeled point in the pool gets a vote from every committee member; the point the committee disagrees about most is the one worth labeling next.
vote entropy: H(x) = -Σ_y (V(y)/K) · log2(V(y)/K)
V(y) = number of committee members voting class y
next query x* = argmax_x H(x) (over the unlabeled pool)
Each committee member is a logistic-regression classifier over features [x, y, x·y, 1], trained by gradient descent on a bootstrap sample (drawn with replacement) of the labeled points — the resampling plus random weight initialization is what makes the K decision boundaries disagree, especially while few labels exist.
- Point size / glow on unlabeled points scales with vote entropy — bigger means the committee is more split, so it's a better query candidate. The outlined point is the current top pick.
- Colored curves are the K committee decision boundaries for the view you've selected; where they fan apart is exactly where disagreement concentrates.
- Two independent experiments run side by side from the same starting labels: one always queries the highest vote-entropy point (QBC), the other queries a uniformly random pool point. Every "Query next" step advances both by one label so the label budget stays matched.
- The accuracy chart plots each experiment's committee accuracy (majority vote over a fixed, noise-free test grid) against labels used — the gap between the two curves is QBC's real, measured sample-efficiency gain.
Real-world relevance: QBC and its relatives (uncertainty sampling, core-set/diversity sampling) are exactly what labeling pipelines at ML teams use to decide which of millions of unlabeled examples are worth sending to a human annotator next.