Research map / Label-efficient and robust learning / Recognition and long-tail

Dynamic Pseudo Labeling via Gradient Cutting for High-Low Entropy Exploration

CVPR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Partial and multi-label learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes gradient-cutting dynamic pseudo-labeling with two gradient-pass classifiers (HGP and LGP) whose local thresholds are combined with the main classifier's threshold to include high-entropy samples in semi-supervised learning.

Problem

Existing dynamic pseudo-labeling methods in semi-supervised image classification rely on confidence thresholds that overlook high-entropy samples and cause overconfidence, leading to poor calibration and limited generalization.

Approach

Two auxiliary classifiers are trained with gradient cutting on the consistency loss: HGP lowers small gradients (encouraging overconfidence) and LGP cuts large gradients (encouraging underconfidence). Their class-wise thresholds are combined via negative-exponential max normalization with the main classifier's threshold to form a gradient-cutting dynamic threshold for pseudo-labeling.

Key results

• On CIFAR-10 with 40 labels the method reaches 4.57% top-1 error versus 4.90% for FreeMatch, and it achieves the lowest average error across benchmarks in Table 3.
• On STL-10 with 40 labels the method reports 13.5% error, compared with 15.56% for FreeMatch and 15.85% for SimMatchV2 (as reported in the table).
• Ablation on CIFAR-10 (40) shows the full combination of T_f, T_HGP, T_LGP (T_GCD) reduces error from 10.04% (T_f alone) to 4.52%.

Measured effects reported in the paper

Surprising findings

• The HGP and LGP auxiliary classifiers have substantially higher error rates than the main classifier, which the authors describe as an expected outcome of their design.

Limitations

• Gradient cutting can cause unstable training for large-scale models.
• The auxiliary GP classifiers substantially increase the training budget.
• Evaluated only on small-scale image classification benchmarks.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 SVHN STL-10

Benchmarks: CIFAR-10 CIFAR-100 SVHN STL-10

Metrics: top-1 classification error rate reliability diagram / ECE

Baselines: PseudoLabeling FixMatch UDA MPL DASH FlexMatch CoMatch AdaMatch SimMatch ConMatch SoftMatch ReFixMatch

Builds on: FreeMatch FixMatch Dash FlexMatch

Closest papers by meaning

• Non-Stationary Predictions May Be More Informative: Exploring Pseudo-Labels with a Two-Phase Pattern of Training Dynamics (ICML 2025 · similarity 0.86)
• Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.85)
• Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning (NeurIPS 2025 · similarity 0.85)
• Semi-Supervised Regression with Heteroscedastic Pseudo-Labels (NeurIPS 2025 · similarity 0.85)
• Decouple then Classify: A Dynamic Multi-view Labeling Strategy with Shared and Specific Information (ICML 2024 · similarity 0.85)
• Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.84)