Abstract:Episodic training is a mainstream training strategy for few-shot learning. In few-shot scenarios, however, this strategy is often inferior to some non-episodic training strategy, e. g., Neighbourhood Component Analysis (NCA), which challenges the principle that training conditions must match testing conditions. Thus, a question is naturally asked: How to search for episodic-free tasks for better few-shot learning? In this work, we propose a novel meta-training framework beyond episodic training. In this framework, episodic tasks are not used directly for training, but for evaluating the effectiveness of some selected episodic-free tasks from a task set that are performed for training the meta-learners. The selection criterion is designed with the affinity, which measures the degree to which loss decreases when executing the target tasks after training with the selected tasks. In experiments, the training task set contains some promising types, e. g., contrastive learning and classification, and the target few-shot tasks are achieved with the nearest centroid classifiers on the miniImageNet, tiered-ImageNet and CIFAR-FS datasets. The experimental results demonstrate the effectiveness of our approach.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: how to select more effective training tasks in few - shot learning (FSL) to surpass the traditional episodic - based training strategies. Specifically, the author focuses on how to design a framework to select better training tasks so that these tasks can improve the performance of the model on the target few - shot tasks. ### Background and Problem Description Few - shot learning aims to solve the problem of classifying a small number of samples, especially when facing unseen classes. Traditional meta - learning frameworks usually adopt episodic training, that is, simulating test conditions by dividing a batch of samples into support sets. However, some studies have shown that in some cases, non - episodic training strategies (such as Neighborhood Component Analysis, NCA) may be superior to episodic training strategies. This challenges the principle that "training conditions must match test conditions" and raises a question: how to find non - episodic - dependent tasks for better few - shot learning? ### Core Contributions of the Paper 1. **Proposing a new meta - training framework**: In this framework, episodic tasks are not directly used for training, but for evaluating the effectiveness of non - episodic tasks selected from the task set. 2. **Introducing affinity as a task - selection criterion**: Affinity measures the degree of loss reduction when performing the target task, thereby selecting tasks with the highest affinity for joint training. 3. **Verifying the effectiveness of the framework through experiments**: The experimental results show that the proposed framework achieves better performance on multiple datasets. ### Method Overview - **Task - selection criterion**: Use affinity to evaluate the impact of different tasks on the target few - shot task. Affinity is defined as: \[ Z_{T_{\text{sub}} \to T_e}^{t, \text{UNA}} = 1 - \frac{\sum_{l = 1}^L L_{\nu_l}(S_l, Q_l; \theta_s^{t+\text{UNA}}, \theta_i)}{\sum_{l = 1}^L L_{\nu_l}(S_l, Q_l; \theta_s^t, \theta_i)} \] where \( Z_{T_{\text{sub}} \to T_e}^{t, \text{UNA}} \) represents the impact of the sub - task set \( T_{\text{sub}} \) on the target task \( T_e \) after the \( t \) - th update. - **Optimization process**: By calculating the average affinity of each sub - task set, select the optimal task set for training: \[ T_{\text{sub}}^*=\arg \min_{T_{\text{sub}}\subset T_{\text{set}}} \hat{Z}_{T_{\text{sub}} \to T_e}^{t, \text{UNA}, M} \] ### Experimental Results The paper verifies the effectiveness of the proposed method through multiple experiments, including hyper - parameter analysis, comparison of different task - selection strategies, and ablation experiments. The results show that the EFTS method outperforms other methods in various settings, especially in 1 - shot and 5 - shot tasks. ### Conclusion By introducing a non - episodic task - selection mechanism, this paper effectively improves the performance of few - shot learning and provides new ideas and methods for future few - shot learning research.

Episodic-free Task Selection for Few-shot Learning

Transductive Episodic-Wise Adaptive Metric for Few-Shot Learning

Task Attended Meta-Learning for Few-Shot Learning

Learning to Learn Adaptive Classifier–Predictor for Few-Shot Learning

Understanding the Overfitting of the Episodic Meta-training

Task-Agnostic Self-Distillation for Few-Shot Action Recognition

Few-shot Learning with Meta Metric Learners

Task-Equivariant Graph Few-shot Learning

Learning Task-aware Local Representations for Few-shot Learning.

Meta-Learning for Semi-Supervised Few-Shot Classification

Learning task-specific discriminative embeddings for few-shot image classification

Few-Shot Classification with Contrastive Learning

Multi-Granularity Episodic Contrastive Learning for Few-Shot Learning

Meta-ticket: Finding optimal subnetworks for few-shot learning within randomly initialized neural networks

Task Cooperation for Semi-Supervised Few-Shot Learning.

SPContrastNet: A Self-Paced Contrastive Learning Model for Few-Shot Text Classification

Episodic fine-tuning prototypical networks for optimization-based few-shot learning: Application to audio classification

The meta-learning method for the ensemble model based on situational meta-task

Few-Shot Learning With Global Class Representations

Few-shot Learning with Adaptively Initialized Task Optimizer: a Practical Meta-Learning Approach

Learning to focus: cascaded feature matching network for few-shot image recognition