Meta and Multi-Task Learning for Action Recognition: A Survey

Human movement recognition currently plays an important role in fields where human-machine interaction is included, such as virtual reality and rehabilitation. In these cases, the accurate recognition of human actions has become essential, and is currently making significant steps due to the wide use of Machine Learning algorithms. However, the standard methods in that category currently face several challenges that concern their ability to generalize and adapt to new, unseen data, as well as the need for large amounts of data for training. This survey examines how the methods of Multi-Task Learning and Meta Learning have been deployed as strategies to tackle the above challenges. It studies how the two approaches are applied to human action recognition, focusing on the utilized methodologies, data modalities, and architectural approaches, highlighting the way that the first method promotes task-related knowledge sharing, while the second enables adaptation to scarce data. Through the mapping of studies from 2014 to 2025, this survey identifies trends, methodological gaps, and future insights for the development of more adaptive and context-aware recognition systems.