Meta-Task: A Method-Agnostic Framework for Learning to Regularize in Few-Shot Learning

TL;DR

Meta-Task presents a method-agnostic regularization framework for few-shot learning by treating regularization as a learnable auxiliary task. It introduces a Task-Decoder autoencoder that refines embeddings through input reconstruction, integrated into existing FSL models with a joint objective $J(\theta)=\sum J_{\mathcal{T}}(\theta)+\lambda\sum J_{\mathcal{M}}(\theta,\phi)$. Empirically, the approach yields faster convergence, higher accuracy, and better generalization across MiniImageNet, TieredImageNet, and FC100 when paired with Prototypical Networks, MAML, MetaOptNet, and $P > M > F$, with modest hyperparameter tuning. This framework has practical impact for resource-constrained scenarios, enabling robust few-shot learning without extensive pretraining or manual loss engineering.

Abstract

Overfitting is a significant challenge in Few-Shot Learning (FSL), where models trained on small, variable datasets tend to memorize rather than generalize to unseen tasks. Regularization is crucial in FSL to prevent overfitting and enhance generalization performance. To address this issue, we introduce Meta-Task, a novel, method-agnostic framework that leverages both labeled and unlabeled data to enhance generalization through auxiliary tasks for regularization. Specifically, Meta-Task introduces a Task-Decoder, which is a simple example of the broader framework that refines hidden representations by reconstructing input images from embeddings, effectively mitigating overfitting. Our framework's method-agnostic design ensures its broad applicability across various FSL settings. We validate Meta-Task's effectiveness on standard benchmarks, including Mini-ImageNet, Tiered-ImageNet, and FC100, where it consistently improves existing state-of-the-art meta-learning techniques, demonstrating superior performance, faster convergence, reduced generalization error, and lower variance-all without extensive hyperparameter tuning. These results underline Meta-Task's practical applicability and efficiency in real-world, resource-constrained scenarios.

Figures (9)

• Figure 1: Detailed architecture of the Task-Decoder within the Meta-Task framework. The Meta-Task support set consists of unlabeled copies of support and query samples, which are fed into an encoder $f_{\theta}$ to produce feature embeddings. These embeddings serve two key functions: (1) they are passed to a decoder $g_{\phi}$ to reconstruct the input images, aiding in the learning of rich, meaningful representations, and (2) they are used to generate the meta-learning method ($F_{\theta}$) for query image classification. The model optimizes two objectives: classification loss ($J_{\mathcal{T}}$) and reconstruction loss ($J_{\mathcal{M}}$), each with a dedicated backpropagation step. The overall training loss is $J_{\mathcal{T}} + \lambda J_{\mathcal{M}}$, balancing both classification accuracy and representation learning. For a broader application, the decoder $g_{\phi}$ could be replaced by one or more neural networks that solve various unsupervised tasks.
• Figure 2: Accuracy curves for Prototypical Networks (PN) and Task-Decoder (TD) across different datasets, plotted against the number of episodes.
• Figure 3: Loss curves for Prototypical Networks (PN) and Task-Decoder (TD) across different datasets, plotted against the number of episodes.
• Figure 4: The reconstruction results of the Prototypical Network integrated with the Task-Decoder, trained on TieredImageNet, are evaluated using MiniImageNet images.
• Figure 5: illustrates the t-SNE projections of the first 10 classes of MiniImageNet for Prototypical Networks with (left) and without (right) the Task-Decoder. The integration of the Task-Decoder demonstrates enhanced class separability, indicating its effectiveness in improving feature representation.

Theorems & Definitions (3)

• Remark 1
• Remark 2
• Remark 3