Research map / Label-efficient and robust learning / Semi-supervised learning
Language-Assisted Debiasing and Smoothing for Foundation Model-Based Semi-Supervised Learning
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 LADaS, a semi-supervised learning framework for foundation models that debiases pseudo-labels with a reliability-scored, language-guided logit adjustment and smooths low-confidence samples with language-aware soft labels.
Problem
Foundation models used in semi-supervised learning are biased toward some classes and produce imbalanced pseudo-labels; existing logit-adjustment debiasing ignores the reliability of its estimated offset and linguistic knowledge, and discards low-confidence unlabeled samples.
Approach
Combines a self-adaptive logit adjustment (EMA class learning pace scaled by a confidence-based reliability score) with a language-oriented preference computed from CLIP-style text class centers and image-text similarity, plus a language-aware soft pseudo-label smoothing loss for low-confidence samples.
Key results
- On CIFAR-100 with ViT-FFT at 2 labels per class, LADaS reaches 18.31% error versus 23.60% for InterLUDE+ (Table 4).
- With EfficientNet-B2 on CIFAR-100 at 4 labels per class, LADaS achieves 41.27% error versus 44.82% for BorLan (Table 3).
- On FOOD-101 with ViT-VPT at 10 labels per class, LADaS reaches 9.77% error, versus 10.31% for FINESSL at the same setting (Table 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LADaS instead of FINESSL | ViT-VPT · FOOD-101 N2 | error rate (%): 12.96 ± 0.43 -> 10.85 ± 0.79 | Helped |
| Use LADaS instead of FINESSL | ViT-VPT · CIFAR-100 N4 | error rate (%): 19.56 ± 0.24 -> 18.75 ± 0.14 | Helped |
| Use LADaS instead of FINESSL | ViT-VPT · CIFAR-10 N4 | error rate (%): 2.61 ± 0.05 -> 2.53 ± 0.08 | Helped |
| Use LADaS instead of FINESSL | ViT-VPT · Semi-Aves Du = Du in ∪ Du out | error rate (%): 38.88 ± 0.07 -> 39.14 ± 0.21 | Helped |
| Use LADaS instead of BorLan | EfficientNet-B2 · CIFAR-100 N4 | error rate (%): 44.82 -> 41.27 | Helped |
| Use LADaS instead of InterLUDE+ | ViT-FFT · CIFAR-10 N4 | error rate (%): 1.55 ± 0.1 -> 1.17± 0.02 | Helped |
| Add logit adjustment based on class learning pace (ablation) | LADaS ablation model · CIFAR-10 N1 | error rate (%): 31.16 -> 21.97 | Helped |
| Add a reliability score for class learning pace (ablation) | LADaS ablation model · CIFAR-10 N1 | error rate (%): 21.97 -> 5.06 | Helped |
| Add language-oriented preference (ablation) | LADaS ablation model · CIFAR-10 N1 | error rate (%): 5.06 -> 3.24 | Helped |
| Add language-aware pseudo-label smoothing loss (ablation) | LADaS ablation model · CIFAR-10 N1 | error rate (%): 4.43 -> 3.15 | Helped |
| Add vanilla label smoothing loss (ablation) | Base model · CIFAR-10 N1 | error rate (%) | Helped |
| Replace vanilla label smoothing with language-aware pseudo-label smoothing (ablation) | Base model · CIFAR-10 N1 | error rate (%) | Helped |
Surprising findings
- On Semi-Aves with OOD unlabeled data, LADaS (39.14%) is not best; FINESSL (38.88%) edges it out, the paper notes LADaS is competitive rather than uniformly superior.
Limitations
- Gains are sometimes marginal or absent; LADaS is not best across all settings (e.g., Semi-Aves with OOD data, some FOOD-101 and CIFAR-10 settings).
- Relies on a frozen PLM (BERT-Large) and 80 handcrafted CLIP-style prompts for class centers, adding dependence on prompt design.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 STL-10 FOOD-101 Semi-Aves
Benchmarks: CIFAR-10 CIFAR-100 STL-10 FOOD-101 Semi-Aves
Metrics: error rate
Baselines: PL FixMatch FlexMatch FreeMatch SoftMatch DebiasPL FINESSL Co-Tuning ReMixMatch Self-Tuning BorLan InstanT
Builds on: FixMatch FreeMatch DebiasPL FINESSL BorLan CLIP
Closest papers by meaning
- Don’t fear the unlabelled: safe semi-supervised learning via debiasing (ICLR 2023 · similarity 0.90)
- Erasing the Bias: Fine-Tuning Foundation Models for Semi-Supervised Learning (ICML 2024 · similarity 0.89)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.87)
- Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL (ICML 2025 · similarity 0.87)
- ABC-Align: Prediction-Powered Alignment with Adaptive Bias Control (NeurIPS 2026 · similarity 0.86)
- Adaptive Integration of Partial Label Learning and Negative Learning for Enhanced Noisy Label Learning (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.