Kim Ihyeon

Self-training (pseudo-labeling) is the simplest and most widely used semi-supervised learning (SSL) method, but when and why it collapses—performance falling far below a supervised-only baseline, not just short of it—remains poorly characterized quantitatively. We study this using protein language model (PLM) embeddings as a controlled testbed, directly manipulating class imbalance on frozen ESM-2 embeddings (DeepLoc 2.0 subcellular localization) to isolate its causal role. Across a fine-grained sweep and multiple seeds per condition, collapse concentrates in a reproducible “danger zone” (roughly 8–20% positive rate), confirmed to reflect training-pool imbalance itself rather than train-test distribution mismatch. Applying seed-level statistical rigor to calibration, we show Expected Calibration Error (ECE) reliably rises whenever self-training collapses (r = −0.91, partial r = −0.92 controlling for imbalance level), while pre-training ECE does not predict later collapse—indicating miscalibration is a necessary but not sufficient precursor. Comparing remedies, calibrated self-training and confidence-threshold tuning both give only partial relief, while explicitly balancing which pseudo-labels are accepted across classes eliminates collapse entirely and slightly exceeds the supervised baseline; masking-based consistency regularization, which never thresholds into hard pseudo-labels, is robust throughout without any calibration correction. We further show the danger zone reproduces at a four-times-larger PLM scale, is not explained by embedding geometry, and persists against binary-adapted DARP and CReST baselines from the class-imbalanced SSL literature—both underperform our class-balanced selection, with DARP collapsing even more than naive self-training despite an oracle advantage unavailable in practice. We frame protein PLMs as a case study for a general SSL failure mode, and release our controlled-imbalance protocol, failure-regime criteria, and remedy comparisons for use in other domains.