Provable Active Multi-Task Representation Learning

Multi-task representation learning is an emerging machine learning paradigm that integrates data from multiple sources, harnessing task similarities to enhance overall model performance. The application of multi-task learning to real-world settings is hindered due to data scarcity, along with challe...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on signal processing S. 1 - 15
Hauptverfasser: Lin, Jiabin, Moothedath, Shana, Le, Tuan Anh
Format: Journal Article
Sprache:Englisch
Veröffentlicht: IEEE 2025
Schlagworte:
active learning
alternating gradient descent and minimization algorithm
Complexity theory
Machine learning algorithms
Minimization
Multi-task representation learning
Multitasking
Noise
Representation learning
Signal processing algorithms
Training
transfer learning
Vectors
ISSN:1053-587X, 1941-0476
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Multi-task representation learning is an emerging machine learning paradigm that integrates data from multiple sources, harnessing task similarities to enhance overall model performance. The application of multi-task learning to real-world settings is hindered due to data scarcity, along with challenges related to scalability and computational resources. To address these challenges, we develop a fast and sample-efficient approach for multi-task active learning with linear representation when the amount of data from source tasks and target tasks is limited. By leveraging the techniques from active learning, we propose an adaptive sampling-based alternating projected gradient descent (GD) and minimization algorithm that iteratively estimates the relevance of each source task to the target task and samples from each source task based on the estimated relevance. We present the convergence guarantees and the sample and time complexities of our algorithm. We evaluated the effectiveness of our algorithm using experiments and compared it with four benchmark algorithms using synthetic and real-world MNIST-C and MovieLens-100K datasets.
ISSN:1053-587X
1941-0476
DOI:10.1109/TSP.2025.3623098