[Paper Review] A Comprehensive Overview and Survey of Recent Advances in Meta-Learning

Huimin Peng|arXiv (Cornell University)|Apr 17, 2020

Domain Adaptation and Few-Shot Learning123 references37 citations

TL;DR

This survey reviews meta-learning methodologies, categorizations, datasets, and applications, with emphasis on few-shot learning and integration with other ML frameworks.

ABSTRACT

This article reviews meta-learning also known as learning-to-learn which seeks rapid and accurate model adaptation to unseen tasks with applications in highly automated AI, few-shot learning, natural language processing and robotics. Unlike deep learning, meta-learning can be applied to few-shot high-dimensional datasets and considers further improving model generalization to unseen tasks. Deep learning is focused upon in-sample prediction and meta-learning concerns model adaptation for out-of-sample prediction. Meta-learning can continually perform self-improvement to achieve highly autonomous AI. Meta-learning may serve as an additional generalization block complementary for original deep learning model. Meta-learning seeks adaptation of machine learning models to unseen tasks which are vastly different from trained tasks. Meta-learning with coevolution between agent and environment provides solutions for complex tasks unsolvable by training from scratch. Meta-learning methodology covers a wide range of great minds and thoughts. We briefly introduce meta-learning methodologies in the following categories: black-box meta-learning, metric-based meta-learning, layered meta-learning and Bayesian meta-learning framework. Recent applications concentrate upon the integration of meta-learning with other machine learning framework to provide feasible integrated problem solutions. We briefly present recent meta-learning advances and discuss potential future research directions.

Motivation & Objective

• Motivate and define meta-learning and its goals for rapid adaptation to unseen tasks.
• Categorize meta-learning approaches into model-based, metric-based, optimization-based, and Bayesian frameworks.
• Summarize datasets and formulations used for few-shot meta-learning and episodic training.
• Outline how meta-learning integrates with reinforcement learning, imitation learning, online learning, and unsupervised learning.
• Highlight future research directions and potential impact on general AI.

Proposed method

• Classify meta-learning methodologies into four broad categories: model-based, metric-based, layered/optimization-based, and Bayesian meta-learning.
• Discuss representative methods such as memory-augmented networks, Siamese/prototypical/matching networks, MAML and its variants, and Bayesian extensions.
• Describe episodic training for few-shot learning and the meta-training/val/testing workflow with support and query sets.
• Review the role of similarity modeling between tasks and how it guides fast adaptation across tasks.
• Summarize applications in meta-RL and meta-imitation learning, as well as unsupervised and online meta-learning.

Experimental results

Research questions

• RQ1 What are the main methodological categories of meta-learning and their core mechanisms?
• RQ2 How do few-shot learning frameworks use episodic training to enable rapid adaptation across tasks?
• RQ3 What datasets and benchmarks are commonly used to evaluate meta-learning approaches, especially for image classification?
• RQ4 How can meta-learning be integrated with other ML paradigms such as reinforcement learning, imitation learning, online learning, and unsupervised learning?
• RQ5 What are the promising directions and challenges identified for future meta-learning research?

Key findings

• Meta-learning methods are broadly categorized into model-based, metric-based, layered/optimization-based, and Bayesian frameworks.
• MAML is highlighted as a prominent, model-agnostic approach applicable to any learner trained with SGD.
• Few-shot learning is the main focus of recent advances, with datasets like miniImageNet and tieredImageNet commonly used for evaluation.
• Meta-learning increasingly integrates with other ML frameworks such as meta-RL, meta-imitation learning, and online/unsupervised meta-learning.
• There is growing emphasis on similarity modeling between tasks to enable generalization across diverse tasks.