Athens University Economics and Business (AUEB)
Federated, continual and meta-learning algorithms for multi-task settings
Abstract
dc:descriptionIn this dissertation, we present three novel Machine Learning algorithms, each addressing Multi-Task scenarios within different contexts. Multi-Task Learning (MTL) involves handling multiple tasks simultaneously or sequentially, where a task is specified in terms of an objective function and a dataset. The number of tasks can be finite or infinite, and these tasks may be distributed across multiple clients or contained within a single client. The algorithms we propose tackle Multi-Task Learning in the contexts of Personalized Federated Learning (PFL), Continual Learning, and Meta-Learning, each optimized for their specific task environments. First, we explore MTL in the context of PFL. Here, MTL is applied to learn a shared representation across a finite set of tasks distributed among multiple clients. This shared representation across clients enable better generalization and allow individual models to learn their tasks more efficiently, particularly when the size of each client’s dataset is limited. In this thesis, we propose a PFL algorithm utilizing exact distributed Stochastic Gradient Descent (SGD) which has theoretical convergence guarantees. Our algorithm handles the finite set of tasks distributed across multiple clients, by jointly optimizing two sets of weights, i.e., common and personalized weights. This method demonstrates superior performance on multi-class classification datasets. Second, we address MTL in the context of CL, where, unlike the standard MTL’s joint learning approach with access to all task data, the learner must adapt to tasks sequentially as they arrive from a stream while retaining good performance across all tasks at all times. We focus on classification problems and task-incremental CL, where new classes are introduced incrementally, requiring the learner to adapt to new classes over time. Unlike MTL, CL does not have access to all available data at once and cannot revisit past data unless it is explicitly stored in memory. This restriction, along with data imbalances inherent in the incoming stream, creates challenges in ensuring fairness in terms of accuracy, as the learner must learn to treat all classes equitably. To address CL’s challenges, we introduce an online replay-based Continual Learning method that minimizes Kullback-Leibler (KL) loss in order to achieve fairness in problems where there is high class imbalance. This approach achieves improved fairness and accuracy across datasets like MNIST, CIFAR-10, and CIFAR-100 compared to state-of-the-art Reservoir Sampling techniques. Finally, we address MTL in the context of Meta-Learning. While MTL enables the simultaneous handling of multiple tasks and promotes the sharing of information across these tasks to enhance generalization and performance, Meta-Learning extends these capabilities by focusing on adapting to an infinite number of tasks drawn from some distribution. A key concept in this context is to model and learn a stochastic representation or encoding along with its distribution so that it captures as much information as possible about each task to make accurate predictions. This is particularly important in environments with distribution drift, such as in CL, where tasks evolve over time. We develop a Meta-Learning framework using mutual information and the information bottleneck principle. This leads to stochastic generalizations of the well known Model Agnostic Meta Learning (MAML) as well as new Memory-based algorithms that apply Gaussian Process kernel methods and achieve competitive performance in few-shot regression and classification.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Nikoloutsopoulos, Sotirios
- Νικολουτσόπουλος, Σωτήριος
Subjects
dc:subject × 16- Ομόσπονδη μάθηση
- Συνεχής μάθηση
- Μετα-μάθηση
- Μηχανική μάθηση
- Περιβάλλοντα πολλαπλών εργασιών
- Federated learning
- Continual learning
- Meta-learning
- Machine learning (ML)
- Multi-task settings
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58349