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

Instance-based Max-margin for Practical Few-shot Recognition

CVPR 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 a practical many-way few-shot setting (pFSL) on unsupervised pre-trained backbones and IbM2, an instance-based max-margin method that trains a linear classifier on noise-augmented virtual features with an ellipsoidal noise scale found by binary search.

Problem

Traditional few-shot learning relies on a base set, uses few-way episodes that are unreliable and costly to evaluate, and is saturated; a more practical setting with many novel classes and no base set is needed, and margin-based classifiers struggle with very few high-dimensional examples.

Approach

Generate virtual samples on hyperspheres around each training feature with ellipsoidal Gaussian noise scaled by per-dimension standard deviations, and binary-search the largest radius for which a linear classifier still classifies the virtual samples correctly, achieving instance-level max-margin without a base set or validation set.

Key results

• On pFSL with ViT-S/16 MSN ImageNet-1K 1-shot to 16-shot, IbM2 improves accuracy at most shots, e.g., 47.9 to 47.8 at 1-shot but 66.1 to 68.4 at 16-shot.
• In traditional 5-way FSL, IbM2 improved mean accuracy ACCm in all but one case, with average ACCm gain of 0.4% and ACC1 gain of 1.3%.
• IbM2 outperformed SVM with best C on ImageNet-1K with ViT-L/7 MSN, e.g., 58.2 vs 54.8 at 1-shot and 76.1 vs 74.7 at 16-shot.

Measured effects reported in the paper

Surprising findings

• Spherical (isotropic) noise sampling slightly degraded accuracy in low-shot cases (<=3 shots) relative to the baseline.
• IbM2 gains are largest for classes of middle difficulty; for already-accurate classes the improvement is small.

Limitations

• Gains at 1-shot are small or absent in several pFSL settings and IbM2 is weakest in the extremely low-shot regime.
• The backbone is frozen and only a linear classifier is learned, so backbone tuning is not explored.
• Relies on a hyperparameter search (binary search with threshold T and R) per task.

Experimental setup

Datasets: ImageNet-1K CUB-200-2011 mini-ImageNet CIFAR-FS

Benchmarks: ImageNet-1K CUB-200-2011 mini-ImageNet CIFAR-FS

Metrics: top-1 accuracy standard deviation ACCm ACC1 ACC10 ACC100

Baselines: SVM Distribution calibration PMF S2M2R Meta-Baseline

Builds on: SVM Distribution calibration Gaussian Annulus Theorem Meta-Baseline PMF

Closest papers by meaning

• Boosting Few-Shot Learning With Adaptive Margin Loss (CVPR 2020 · similarity 0.85)
• Few-Shot Recognition via Stage-Wise Retrieval-Augmented Finetuning (CVPR 2025 · similarity 0.83)
• FSL-Rectifier: Rectify Outliers in Few-Shot Learning via Test-Time Augmentation (AAAI 2025 · similarity 0.82)
• Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.82)
• Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.82)
• Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.82)