Meta-Learning by Adjusting Priors Based on Extended PAC-Bayes Theory

Abstract

In meta-learning an agent extracts knowledge from observed tasks, aiming to facilitate learning of novel future tasks. Under the assumption that future tasks are 'related' to previous tasks, the accumulated knowledge should be learned in a way which captures the common structure across learned tasks, while allowing the learner sufficient flexibility to adapt to novel aspects of new tasks. We present a framework for meta-learning that is based on generalization error bounds, allowing us to extend various PAC-Bayes bounds to meta-learning. Learning takes place through the construction of a distribution over hypotheses based on the observed tasks, and its utilization for learning a new task. Thus, prior knowledge is incorporated through setting an experience-dependent prior for novel tasks. We develop a gradient-based algorithm which minimizes an objective function derived from the bounds and demonstrate its effectiveness numerically with deep neural networks. In addition to establishing the improved performance available through meta-learning, we demonstrate the intuitive way by which prior information is manifested at different levels of the network.

Accepted to ICML 2018

Cited by in corpus (34)

• A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges
• Generalizing from a Few Examples: A Survey on Few-Shot Learning
• Learning to Demodulate from Few Pilots via Offline and Online Meta-Learning
• Empirical Bayes Transductive Meta-Learning with Synthetic Gradients
• Meta-Learning without Memorization
• From Learning to Meta-Learning: Reduced Training Overhead and Complexity for Communication Systems
• User-friendly introduction to PAC-Bayes bounds
• A Primer on PAC-Bayesian Learning
• Invariant Causal Prediction for Block MDPs
• Provable Guarantees for Gradient-Based Meta-Learning
• All You Need is a Good Functional Prior for Bayesian Deep Learning
• Adaptive Gradient-Based Meta-Learning Methods
• Meta Discovery: Learning to Discover Novel Classes given Very Limited Data
• Generalization Bounds For Meta-Learning: An Information-Theoretic Analysis
• A Theoretical Analysis of the Number of Shots in Few-Shot Learning
• Meta-learners' learning dynamics are unlike learners'
• AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning
• PAC-Bayes Bounds for Meta-learning with Data-Dependent Prior
• Learning Robust State Abstractions for Hidden-Parameter Block MDPs
• How Fine-Tuning Allows for Effective Meta-Learning
• Efficient hyperparameter optimization by way of PAC-Bayes bound minimization
• Information-Theoretic Analysis of Epistemic Uncertainty in Bayesian Meta-learning
• Upper and Lower Bounds on the Performance of Kernel PCA
• PAC-Bayes Analysis of Sentence Representation
• Theoretical bounds on estimation error for meta-learning
• How Tight Can PAC-Bayes be in the Small Data Regime?
• A PAC-Bayesian Perspective on Structured Prediction with Implicit Loss Embeddings
• Exploiting a Zoo of Checkpoints for Unseen Tasks
• Margin-Based Transfer Bounds for Meta Learning with Deep Feature Embedding
• Meta R-CNN : Towards General Solver for Instance-level Few-shot Learning
• Metalearning Linear Bandits by Prior Update
• Generalization Bounds for Meta-Learning via PAC-Bayes and Uniform Stability
• Transfer Bayesian Meta-learning via Weighted Free Energy Minimization
• A method of supervised learning from conflicting data with hidden contexts