Meta-learning, or learning to learn, is a rapidly evolving area in machine learning that aims to enhance the adaptability and efficiency of learning algorithms. Inspired by the human ability to learn new concepts from limited examples and quickly adapt to unforeseen situations, meta-learning leverages prior experience to prepare models for fast adaptation to new tasks. Unlike traditional machine learning systems, where models are trained for specific tasks, meta-learning frameworks enable models to acquire generalized knowledge during training and efficiently learn new tasks during inference. This ability to generalize from past experiences to new tasks makes meta-learning a key focus in advancing artificial intelligence, offering the potential to create more flexible and efficient AI systems capable of performing well with minimal data.

In this thesis, we begin by formally defining the meta-learning framework, establishing clear terminology, and synthesizing existing work in a comprehensive survey paper. Building on this foundation, we demonstrate how meta-learning can be integrated into various fields to enhance model performance and extend capabilities to few-shot learning scenarios. We show how meta-learning can significantly improve the accuracy and efficiency of transferring knowledge across domains in domain adaptation. In scenarios involving a multimodal distribution of tasks, we develop methods that efficiently learn from and adapt to a wide variety of tasks drawn from different modes within the distribution, ensuring effective adaptation across diverse domains. Our work on personalized federated learning highlights meta-learning's potential to tailor federated learning processes to individual user needs while maintaining privacy and data security. Additionally, we address the challenges of continual learning by developing models that continuously integrate new information without forgetting previously acquired knowledge. For time series data analysis, we present meta-learning strategies that automatically learn optimal augmentation techniques, enhancing model predictions and offering robust solutions for real-world applications. Lastly, our pioneering research on unsupervised meta-learning via in-context learning explores innovative approaches for constructing tasks and learning effectively from unlabeled data.

Overall, the contributions of this thesis emphasize the potential of meta-learning techniques to improve performance across diverse research areas and demonstrate how advancements in one area can benefit the field as a whole.

This paper introduces a novel transfer learning adapter, the Bridged Attention Module (BAM), designed to enhance the performance of Spatial-Temporal Graph Convolutional Networks (ST-GCN) in data-limited forecasting scenarios. BAM improves fine-tuning efficiency by jointly capturing spatial and temporal dependencies, optimizing information flow, and significantly reducing the number of trainable parameters while preserving model accuracy. Experimental evaluations demonstrate that the BAM-enhanced ST-GCN consistently achieves competitive accuracy and, in some cases, surpasses traditional fine-tuning methods, even with limited data. The effectiveness of this approach is validated using electric vehicle (EV) charging station occupancy forecasting, highlighting the practical utility of BAM. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

In the field of Human Activity Recognition (HAR), the rapid evolution of wearable devices necessitates models that are generalizable and can adapt to entirely new subjects and activities with very limited labeled data. Conventional deep learning models, constrained by their reliance on large training datasets and limited adaptability to novel scenarios, face challenges in these settings. This paper introduces a novel few-shot HAR strategy employing meta-learning, which facilitates rapid adaptation to unseen subjects and activities using minimal annotated samples. Our method augments time series data with a range of transformations, each assigned a learnable weight, enabling the model to prioritize the most effective augmentations and discard the irrelevant ones. Throughout the meta-training phase, the model learns to identify an optimally weighted combination of these transformations, significantly improving the model’s adaptability and generalization to new situations with scarce labeled data. During meta-testing, this knowledge enables the model to efficiently learn from and adapt to a very limited set of labeled samples from completely new subjects undertaking entirely new activities. Extensive experiments on various HAR datasets demonstrate our method’s enhanced adaptability and generalization to tasks never encountered during training, achieving a performance improvement of up to  across all datasets. These results affirm MADA’s potential for real-world applications characterized by limited data availability. The Author(s) 2025

Meta-learning empowers learning systems with the ability to acquire knowledge from multiple tasks, enabling faster adaptation and generalization to new tasks. This review provides a comprehensive technical overview of meta-learning, emphasizing its importance in real-world applications where data may be scarce or expensive to obtain. The paper covers the state-of-the-art meta-learning approaches and explores the relationship between meta-learning and multi-task learning, transfer learning, domain adaptation and generalization, selfsupervised learning, personalized federated learning, and continual learning. By highlighting the synergies between these topics and the field of meta-learning, the paper demonstrates how advancements in one area can benefit the field as a whole, while avoiding unnecessary duplication of efforts. Additionally, the paper delves into advanced meta-learning topics such as learning from complex multi-modal task distributions, unsupervised metalearning, learning to efficiently adapt to data distribution shifts, and continual meta-learning. Lastly, the paper highlights open problems and challenges for future research in the field. By synthesizing the latest research developments, this paper provides a thorough understanding of meta-learning and its potential impact on various machine learning applications. We believe that this technical overview will contribute to the advancement of meta-learning and its practical implications in addressing realworld problems.

Continual learning (CL) refers to the ability to continually learn over time by accommodating new knowledge while retaining previously learned experience. While this concept is inherent in human learning, current machine learning methods are highly prone to overwrite previously learned patterns and thus forget past experience. Instead, model parameters should be updated selectively and carefully, avoiding unnecessary forgetting while optimally leveraging previously learned patterns to accelerate future learning. Since hand-crafting effective update mechanisms is difficult, we propose meta-learning a transformer-based optimizer to enhance CL. This meta-learned optimizer uses attention to learn the complex relationships between model parameters across a stream of tasks, and is designed to generate effective weight updates for the current task while preventing catastrophic forgetting on previously encountered tasks. Evaluations on benchmark datasets like SplitMNIST, RotatedMNIST, and SplitCIFAR-100 affirm the efficacy of the proposed approach in terms of both forward and backward transfer, even on small sets of labeled data, highlighting the advantages of integrating a meta-learned optimizer within the continual learning framework. 

The standard machine learning assumption that training and test data are drawn from the same probability distribution does not hold in many real-world applications due to the inability to reproduce testing conditions at training time. Existing unsupervised domain adaption (UDA) methods address this problem by learning a domain-invariant feature space that performs well on available source domain(s) (labeled training data) and the specific target domain (unlabeled test data). In contrast, instead of simply adapting to domains, this paper aims for an approach that learns to adapt effectively to new unlabeled domains. To do so, we leverage meta-learning to optimize a neural network such that an unlabeled adaptation of its parameters to any domain would yield a good generalization on this latter. The experimental evaluation shows that the proposed approach outperforms standard approaches even when a small amount of unlabeled test data is used for adaptation, demonstrating the benefit of meta-learning prior knowledge from various domains to solve UDA problems.

Few-shot meta-learning involves training a model on multiple tasks to enable it to efficiently adapt to new, previously unseen tasks with only a limited number of samples. However, current meta-learning methods assume that all tasks are closely related and belong to a common domain, whereas in practice, tasks can be highly diverse and originate from multiple domains, resulting in a multimodal task distribution. This poses a challenge for existing methods as they struggle to learn a shared representation that can be easily adapted to all tasks within the distribution. To address this challenge, we propose a meta-learning framework that can handle multimodal task distributions by conditioning the model on the current task, resulting in a faster adaptation. Our proposed method learns to encode each task and generate task embeddings that modulate the model’s activations. The resulting modulated model becomes specialized for the current task and leads to more effective adaptation. Our framework is designed to work in a realistic setting where the mode from which a task is sampled is unknown. Nonetheless, we also explore the possibility of incorporating auxiliary information, such as the task-mode-label, to further enhance the performance of our method if such information is available. We evaluate our proposed framework on various few-shot regression and image classification tasks, demonstrating its superiority over other state-of-the-art meta-learning methods. The results highlight the benefits of learning to embed task-specific information in the model to guide the adaptation when tasks are sampled from a multimodal distribution. © The Author(s) 2024.

Federated learning has emerged as a promising approach for training machine learning models on decentralized data sources while preserving data privacy. However, challenges such as communication bottlenecks, heterogeneity of client devices, and non-i.i.d. data distribution pose significant obstacles to achieving optimal model performance. We propose a novel framework that combines federated learning with meta-learning techniques to enhance both efficiency and generalization capabilities. Our approach introduces a federated modulator that learns contextual information from data batches and uses this knowledge to generate modulation parameters. These parameters dynamically adjust the activations of a base model, which operates using a MAML-based approach for model personalization. Experimental results across diverse datasets highlight the improvements in convergence speed and model performance compared to existing federated learning approaches. These findings highlight the potential of incorporating contextual information and meta-learning techniques into federated learning, paving the way for advancements in distributed machine learning paradigms. Copyright © 2024 by SIAM.

Meta-learning or learning to learn involves training a model on various learning tasks in a way that allows it to quickly learn new tasks from the same distribution using only a small amount of training data (i.e., few-shot learning). Current meta-learning methods implicitly assume that the distribution over tasks is unimodal and consists of tasks belonging to a common domain, which significantly reduces the variety of task distributions they can handle. However, in real-world applications, tasks are often very diverse and come from multiple different domains, making it challenging to meta-learn common knowledge shared across the entire task distribution. In this paper, we propose a method for meta-learning from a multimodal task distribution. The proposed method learns multiple sets of meta-parameters (acting as different initializations of a neural network model) and uses a task encoder to select the best initialization to fine-tune for a new task. More specifically, with a few training examples from a task sampled from an unknown mode, the proposed method predicts which set of meta-parameters (i.e., model’s initialization) would lead to a fast adaptation and a good post-adaptation performance on that task. We evaluate the proposed method on a diverse set of few-shot regression and image classification tasks. The results demonstrate the superiority of the proposed method compared to other state of-the-art meta-learning methods and the benefit of learning multiple model initializations when tasks are sampled from a multimodal task distribution. © 2023 IEEE.

In the field of Human Activity Recognition (HAR), the rapid evolution of wearable devices necessitates models that are generalizable and can adapt to entirely new subjects and activities with very limited labeled data. Conventional deep learning models, constrained by their reliance on large training datasets and limited adaptability to novel scenarios, face challenges in these settings. This paper introduces a novel few-shot HAR strategy employing meta-learning, which facilitates rapid adaptation to unseen subjects and activities using minimal annotated samples. Our approach augments time series data with a range of transformations, each assigned a learnable weight, enabling the model to prioritize the most effective augmentations and discard the irrelevant ones. Throughout the meta-training phase, the model learns to identify an optimal weighted combination of these transformations, significantly improving the model's adaptability and generalization to new situations with scarce labeled data. During meta-testing, this knowledge enables the model to efficiently learn from and adapt to a very limited set of labeled samples from completely new subjects undertaking entirely new activities. Extensive experiments on various HAR datasets demonstrate our method's enhanced adaptability and generalization to tasks never encountered during training, affirming its potential for real-world applications characterized by limited data availability.