Confidence-Calibrated Consistency Matching for Semi-Supervised Image Classification Under Extreme Label Scarcity
TL;DR
This work proposes CCM (Confidence-Calibrated Consistency Matching)—a per-class curriculum threshold, a dual strong-view consistency loss, and a smooth confidence weighting that softly admits borderline pseudo-labels—together with a unified view in which FixMatch, FlexMatch, SoftMatch, and CCM instantiate a single generalized weighting function.
Abstract
Labeling images is expensive, but unlabeled data is abundant. Semi-supervised learning (SSL) addresses this gap, though the dominant pseudo-labeling methods can suffer from confirmation bias—reinforcing their own confident-but-wrong predictions—most severely when labels are scarcest. Under a controlled, reproducible compute-constrained protocol on CIFAR-10, SVHN, and CIFAR-100, we examine which ingredients of consistency-based SSL actually help when as few as four labels per class are available. We propose CCM (Confidence-Calibrated Consistency Matching)—a per-class curriculum threshold, a dual strong-view consistency loss, and a smooth confidence weighting that softly admits borderline pseudo-labels—together with a unified view in which FixMatch, FlexMatch, SoftMatch, and CCM instantiate a single generalized weighting function. On CIFAR-10 with 40 labels, CCM reaches 35.14%, a significant improvement over the FlexMatch design it directly extends (+3.30 percentage points (pp), paired t-test p = 0.001). SoftMatch, re-trained under the identical budget, performs better still at the two smallest budgets (37.04% at 40 labels, p = 0.049), while CCM leads numerically at 4000 labels: the two smooth-weighting designs top the extreme-scarcity board—convergent evidence that the smoothness of the weighting function, more than the placement of its threshold, is the decisive design axis. We also report a negative result: cross-view agreement helps neither as an admission gate nor as reliability reweighting, reducing accuracy by up to 4.37 pp; agreement is a positive correctness signal, but its absolute level (approximately 44% correct among agreeing pseudo-labels) is too low to filter on safely. Curriculum thresholding, by contrast, hurts on the easier SVHN dataset and fails outright on CIFAR-100 when its per-class statistics become too thin. CCM adds no inference-time cost. We do not claim universality; we characterize when each ingredient helps within a single, identical-budget protocol.