自主机器学习

Conventional machine learning (ML) is capable of training on data and constructing models for given tasks, enabling these models to acquire predictive and decision-making abilities for the corresponding tasks. This can be summarized as task-driven ML. Existing methods rely heavily on external human guidance and empirical specification of data and tasks, model configuration, and parameter learning during the learning process. In real world, dynamic, and open environments, these methods cannot learn autonomously as humans do. This paper introduces a new concept of autonomous ML. Specifically, we define autonomous ML as a self-driven, dynamically self-evolving learning process that encompasses self-optimization and self-evolution. Firstly, it can actively explore and perceive the environment for autonomous data selection, autonomous model adaptation, and task switching without human intervention. Simultaneously, autonomous ML is able to dynamically self-evolve based on feedback from environmental learning and the current machine state. Additionally, the paper presents several application case studies of autonomous ML and discusses future research directions. We anticipate that autonomous ML can enable machines to learn autonomously like humans and provide a new perspective for general artificial intelligence.