Source-linked AI summary
Transductive Episodic-Wise Adaptive Metric for Few-Shot Learning
Limeng Qiao, Yemin Shi, Jia Li, Yaowei Wang, Tiejun Huang, Yonghong Tian
TL;DR
Few-shot learning seeks generalizable classifiers from extremely limited labeled data, but adaptation to each task remains challenging. TEAM combines shared embedding, transductive episodic metric adaptation via an SDP with a closed-form on-the-fly solution, and bi-directional similarity; experiments on three benchmark datasets report state-of-the-art performance.
Problem
Few-shot classification needs generalizable classifiers that adapt to specific tasks despite severely limited labeled data.
Method
TEAM learns a shared task-agnostic embedding, constructs each task's metric from pairwise constraints and a regularization prior via an SDP, and uses bi-directional similarity for query–prototype relationships.
Results
TEAM achieves state-of-the-art performance on three few-shot benchmark datasets.
Takeaways & Limitations
TEAM adapts shared features into more discriminative task-specific metric spaces and is reported as easily extendable to semi-supervised learning.
Takeaways & Limitations
The choice of prior metric M0 influences generalization, although the experiments use the identity matrix M0 = I.
Abstract
from arXiv · showhide
Few-shot learning, which aims at extracting new concepts rapidly from extremely few examples of novel classes, has been featured into the meta-learning paradigm recently. Yet, the key challenge of how to learn a generalizable classifier with the capability of adapting to specific tasks with severely limited data still remains in this domain. To this end, we propose a Transductive Episodic-wise Adaptive Metric (TEAM) framework for few-shot learning, by integrating the meta-learning paradigm with both deep metric learning and transductive inference. With exploring the pairwise constraints and regularization prior within each task, we explicitly formulate the adaptation procedure into a standard semi-definite programming problem. By solving the problem with its closed-form solution on the fly with the setup of transduction, our approach efficiently tailors an episodic-wise metric for each task to adapt all features from a shared task-agnostic embedding space into a more discriminative task-specific metric space. Moreover, we further leverage an attention-based bi-directional similarity strategy for extracting the more robust relationship between queries and prototypes. Extensive experiments on three benchmark datasets show that our framework is superior to other existing approaches and achieves the state-of-the-art performance in the few-shot literature.
1. Introduction
Few-shot classification addresses low-data settings where deep networks can overfit, while existing meta-learning approaches may overlook task-specific information and support–query interactions. TEAM combines task adaptation with transductive inference to construct episodic-wise metrics and improve query classification.
- Deep networks can overfit and severely collapse in low-data scenarios when only rare labeled samples are available.
- Meta-learning uses episodic training over independent few-shot tasks to learn classifiers that generalize to unseen target classification problems.
- Existing methods embed examples from different tasks into one task-independent metric space, neglecting distinctive task-level information.
- Inductive inference predicts queries individually and does not adequately exploit interactions between support sets and unlabeled test sets.
- TEAM learns a shared task-agnostic embedding while constructing a task-specific metric from pairwise constraints and a regularization prior through an SDP formulation.
- TEAM additionally introduces transductive inference and a bi-directional similarity strategy, with experiments on three benchmark datasets reporting superiority over state-of-the-art approaches.
2. Related Work
Related work spans meta-learning, task-adaptation methods, and transductive inference for few-shot learning. TEAM differs by constructing task-specific metrics on the fly and using bi-directional query–prototype similarity.
- Meta-learning in Few-shot Learning: Meta-learning learns from performance across tasks to support faster adaptation to new tasks, including parameter updates, initialization, and gradient-based strategies.
- Meta-learning in Few-shot Learning: Several meta-learning approaches suffer from fine-tuning, whereas TEAM solves unseen target tasks explicitly for each task using pairwise constraints and a regularization prior.
- TEAM framework: TEAM constructs a task-specific metric after shared feature extraction and assigns query labels using bi-directional similarity in the adapted space.
- Task Adaptation Approaches: Task-adaptation methods such as MT-Net and TADAM incorporate task identity through activation subspaces or task-dependent scaling and shift vectors.
- Transductive Inference: Transductive inference generalizes directly from training to test data and can improve performance over inductive methods in data-scarce settings.
3. Transductive Episodic-wise Adaptive Metric
TEAM organizes few-shot learning episodically, adapting a shared task-agnostic embedding into a task-specific metric while using the query set transductively. It combines pairwise constraints, regularization, efficient metric optimization, and bi-directional query–prototype similarity.
- Framework: TEAM trains a task-agnostic feature extractor episodically and adapts its embeddings into an episodic-wise task-specific metric.The framework separates shared representation learning from task-level metric construction.
- Framework: Transductive adaptation treats the query set as a whole to construct a more generalizable task-adaptive metric from scarce support data.Queries contribute jointly rather than being processed one by one.
- Episodic-wise Adaptive Metric: Pair-constrained loss minimizes distances for similar pairs while requiring dissimilar pairs to remain farther apart, and regularization keeps the metric close to a prior.The metric uses must-link and cannot-link constraints while controlling over-fitting caused by few available pairs.
- Episodic-wise Adaptive Metric: The similar and dissimilar constraints combine support samples, prototypes, query neighbors, and prototypes from seen classes.Prototype construction shrinks each support set before defining these task-level constraints.
- Episodic-wise Adaptive Metric: The adaptation objective is a semidefinite program, but TEAM derives a closed-form solution using matrix operations that are more efficient than iterative optimizers or naive SDP solvers.The metric remains positive semidefinite, and a positive definite prior supports the stated assumption for the solution.
- Bi-directional Similarity: Bi-SIM multiplies query-to-prototype similarity by prototype-to-query similarity, with the reverse direction acting as an attention-based weight over the query set.This strategy is intended to extract a more robust relationship between queries and prototypes.
4. Experiments
Experiments evaluate TEAM on three few-shot benchmarks using standard 5-way 1-shot and 5-way 5-shot settings, alongside transductive, ablation, semi-supervised, and sparsity analyses. TEAM consistently improves few-shot performance, with especially strong gains in 1-shot settings, while its adaptive metric exhibits sparse, predominantly diagonal structure.
- Experimental Settings: TEAM is evaluated on miniImageNet, Cifar-100, and CUB using 5-way 1-shot and 5-way 5-shot tasks with 15 queries per episode.The experiments also include transductive settings across all datasets.
- Few-shot Learning Results: 51.68% and 68.71% are ProtoNet accuracies on miniImageNet with ConvNet for 5-way 1-shot and 5-way 5-shot, respectively.TEAM improves over the published state-of-the-art by 1.06% and 2.18%, and over the re-implemented baseline by 4.89% and 3.33%.
- Few-shot Learning Results: TEAM consistently improves few-shot performance across datasets and backbones, with larger promotion in 1-shot than 5-shot scenarios.As the number of shots increases, TEAM continues to outperform the baseline, but the improvement decreases slightly.
- Ablation Study: 2.38% is the miniImageNet 1-shot improvement from adding EAM to TIM over the TIM-only variant, demonstrating the value of episodic-wise metric adaptation.The full TEAM configuration additionally combines TIM, EAM, and Bi-SIM.
- Metric Analysis: TEAM’s episodic-wise adaptive metric has larger diagonal than off-diagonal values, with a large gap between them in an oracle 5-way 5-shot analysis.The observed sparsity is attributed to limited data supporting strong self-correlation but insufficient prior for accurate cross-dimension correlations.
5. Conclusions
TEAM combines a shared embedding model with an episodic-wise metric tailored to information within each task, while using the entire query set and bi-directional similarity for inference. It achieves state-of-the-art performance on three few-shot benchmark datasets and can be extended to semi-supervised learning.
- TEAM learns a shared embedding model across tasks and tailors an episodic-wise metric using distinctive task-specific information.
- Using the entire query set at once, TEAM applies bi-directional similarity to extract more robust relationships between queries and prototypes.
- TEAM achieves state-of-the-art performance on three few-shot benchmark datasets and is easily extended to a semi-supervised version.