A PAC-Bayesian bound for Lifelong Learning

Abstract

Transfer learning has received a lot of attention in the machine learning community over the last years, and several effective algorithms have been developed. However, relatively little is known about their theoretical properties, especially in the setting of lifelong learning, where the goal is to transfer information to tasks for which no data have been observed so far. In this work we study lifelong learning from a theoretical perspective. Our main result is a PAC-Bayesian generalization bound that offers a unified view on existing paradigms for transfer learning, such as the transfer of parameters or the transfer of low-dimensional representations. We also use the bound to derive two principled lifelong learning algorithms, and we show that these yield results comparable with existing methods.

to appear at ICML 2014

References in corpus (1)

• Learning Task Grouping and Overlap in Multi-task Learning

Cited by in corpus (47)

• A continual learning survey: Defying forgetting in classification tasks
• Recent Advances in Open Set Recognition: A Survey
• Transductive Multi-view Zero-Shot Learning
• Domain Generalization by Marginal Transfer Learning
• The Benefit of Multitask Representation Learning
• Meta-Learning without Memorization
• From Learning to Meta-Learning: Reduced Training Overhead and Complexity for Communication Systems
• User-friendly introduction to PAC-Bayes bounds
• Recent Advances in Zero-shot Recognition
• Provable Guarantees for Gradient-Based Meta-Learning
• From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence
• Adaptive Gradient-Based Meta-Learning Methods
• PAC-Bayes and Domain Adaptation
• Generalization Bounds For Meta-Learning: An Information-Theoretic Analysis
• PAC-Bayesian Theory Meets Bayesian Inference
• Continual Learning in Neural Networks
• Lifelong Person Re-Identification via Adaptive Knowledge Accumulation
• Transductive Multi-label Zero-shot Learning
• AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning
• Continual Representation Learning for Biometric Identification
• PAC-Bayesian Theorems for Domain Adaptation with Specialization to Linear Classifiers
• PAC-Bayes Bounds for Meta-learning with Data-Dependent Prior
• PAC-Bayes Analysis Beyond the Usual Bounds
• How Fine-Tuning Allows for Effective Meta-Learning
• Information-Theoretic Analysis of Epistemic Uncertainty in Bayesian Meta-learning
• On Localized Discrepancy for Domain Adaptation
• Layerwise Optimization by Gradient Decomposition for Continual Learning
• Bayes meets Bernstein at the Meta Level: an Analysis of Fast Rates in Meta-Learning with PAC-Bayes
• An Information-Geometric Distance on the Space of Tasks
• Conditional Meta-Learning of Linear Representations
• PAC-Bayes Bounds for Bandit Problems: A Survey and Experimental Comparison
• The Advantage of Conditional Meta-Learning for Biased Regularization and Fine-Tuning
• PAC-Bayes Analysis of Sentence Representation
• Theoretical bounds on estimation error for meta-learning
• Risk and Regret of Hierarchical Bayesian Learners
• Unifying Variational Inference and PAC-Bayes for Supervised Learning that Scales
• A Distribution-Dependent Analysis of Meta-Learning
• An Adaptive Online HDP-HMM for Segmentation and Classification of Sequential Data
• On regret bounds for continual single-index learning
• Generalization Bounds for Meta-Learning via PAC-Bayes and Uniform Stability
• Learning Mixtures of Low-Rank Models
• A method of supervised learning from conflicting data with hidden contexts
• Transfer Bayesian Meta-learning via Weighted Free Energy Minimization
• Margin-Based Transfer Bounds for Meta Learning with Deep Feature Embedding
• Accumulating Knowledge for Lifelong Online Learning
• Online Parameter-Free Learning of Multiple Low Variance Tasks
• Metalearning Linear Bandits by Prior Update