Sharing to learn and learning to share : Fitting together metalearning and multi-task learning

Sammanfattning: All humans are born with the ability to learn. The learning process in the brain is innate and natural. Even before birth, it is capable of learning and remembering. As a consequence, humans do not learn everything from scratch, and because they arenaturally capable of effortlessly transferring their knowledge between tasks, they quickly learn new skills. Humans have a natural tendency to believe that similar tasks have(somewhat) similar solutions or approaches, so sharing knowledge from a previous activity makes it feasible to learn a new task easily in a few tries. For instance, the skills learned when riding a bike or a motorbike are helpful while learning to drive a car. Thisnatural learning process, which involves sharing information between tasks, has inspired Deep Learning (DL) to create similar neurally-weighted algorithms. Transfer learning, Multi-task Learning (MTL), meta learning, Lifelong Learning (LL) and many more, areexamples of information sharing algorithms, which exploit the knowledge gained fromone task to improve the performance of another related task. But they vary in terms of what information they share, when to share and why to share.This thesis focuses on two such information sharing algorithms, that are Multi-taskLearning and meta learning. MTL involves learning a number of tasks simultaneously within a shared network structure such that the tasks can mutually benefit each other’s learning. While meta learning better known as ‘Learning to learn’, is an approach ofreducing the amount of time and computation required to learn a novel task by leveraging on knowledge accumulated over the course of numerous training episodes of various tasks. This monograph presents a comprehensive explanation of both meta learning and MTL. A theoretical comparison of both algorithms demonstrates that the strengths of one can outweigh the constraints of the other. Therefore, the objective of this work is to combine MTL and meta learning to attain the best of both worlds.This thesis proposes Multi-task Meta Learning (MTML), an integration of MTL and meta learning. It uses MTL’s ability to train multiple tasks at the same time and meta learning’s ability to quickly and better adapt to a new task. The basic idea is to train a multi-task model so that when a new task is added, it can learn in fewer steps and perform at least as good as traditional single-task learning on the new task or when added to the MTL. The MTML paradigm is demonstrated on two publicly available datasets – theNYU-v2 and the taskonomy dataset, for which four tasks are considered i.e., semantic segmentation, depth estimation, surface normal estimation, and edge detection. This work presents a comparative empirical analysis of MTML to single task and multi-task learning, where MTML excels for most of the task.The future direction of this work includes developing efficient and autonomous MTL architectures by exploiting the concepts of meta learning. The main goal will be tocreate a task adaptive MTL, where meta learning may learn to select layers (or features)from the shared structure for every task because not all tasks require the same high level fine grained features from the shared network. This can be seen as another way of combining MTL and meta learning. It will also introduce modular learning in themulti-task architecture. Furthermore, this work can be extended to include multi-modal multi-task learning, which will be useful to study the contributions of each input modality to various tasks.