Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

BECLR: Batch Enhanced Contrastive Few-Shot Learning

ICLR 2024 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Alignment-based few-shot metric learning · 2 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 BECLR, a contrastive unsupervised few-shot learning method with a dynamic clustered memory (DyCE) for enhanced positive sampling and an optimal-transport distribution alignment (OpTA) at inference to reduce support-query sample bias.

Problem

Unsupervised few-shot learning suffers from instance-level contrastive positives being wrongly repelled within batches and from sample bias between few labeled support prototypes and the unlabeled query distribution at inference.

Approach

A dynamically updated clustered memory with optimal-transport-based equipartitioned prototypes supplies extra nearest-neighbor positives for contrastive pretraining (DyCE), and an iterative optimal-transport alignment moves support prototypes toward query embeddings at inference without learnable parameters (OpTA).

Key results

• BECLR with ResNet-50 reaches 80.57% (1-shot) and 87.82% (5-shot) on miniImageNet, versus 66.68% and 85.29% for the best prior U-FSL baseline (Meta-DM + UniSiam + dist).
• With ResNet-18, BECLR reports 75.74% (1-shot) and 84.93% (5-shot) on miniImageNet, and 76.44% / 84.85% on tieredImageNet.
• Ablation on miniImageNet 5-way 1-shot: adding OpTA gives +12.8% over the DyCE-only configuration (67.75% to 80.57%).

Measured effects reported in the paper

Surprising findings

• OpTA degrades performance when the pretrained encoder is weak, and only gives increasing gains as pretraining accuracy improves.
• The negative (repulsion) loss term is unnecessary to prevent collapse; lambda=0 performed comparably to lambda=0.1 in the hyperparameter ablation (85.45 vs 87.82 on miniImageNet 5-shot).

Limitations

• OpTA can hurt performance when the pretrained feature quality is poor.
• Memory size M and cluster count P are dataset-dependent hyperparameters that need tuning on an unlabeled training set.

Experimental setup

Datasets: miniImageNet tieredImageNet CIFAR-FS FC100 CUB ChestX ISIC EuroSAT CropDiseases

Benchmarks: miniImageNet tieredImageNet CIFAR-FS FC100 CUB ChestX ISIC EuroSAT CropDiseases

Metrics: accuracy (%) with 95% confidence intervals over 2000 episode

Baselines: SwAV NNCLR SimSiam PsCo UniSiam + dist Meta-DM + UniSiam + dist SAMPTransfer HMS PDA-Net ConFeSS SimCLR MoCo v2

Builds on: NNCLR UniSiam PsCo Sinkhorn-Knopp

Closest papers by meaning

• Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.88)
• OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning (CVPR 2024 · similarity 0.88)
• Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.88)
• $\mathbb{X}$-Sample Contrastive Loss: Improving Contrastive Learning with Sample Similarity Graphs (ICLR 2025 · similarity 0.88)
• Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)
• Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.87)