Knowledge transfer and training strategies to train multiple tasks with a lifelong learning network

Abstract

With the development of deep learning, the tasks that Artificial Intelligence (AI) can solve with capabilities beyond humans have diversified, such as computer vision, natural language processing, and robotics. Here, since all tasks are trained under the common assumption that the distribution of training and test data is the same, the neural network trained for one task cannot perform other tasks. However, in the real world, AI must respond to multiple tasks or data domains that appear simultaneously or sequentially. This paper defines three cases in which one neural network can continuously exert knowledge while training multiple tasks and presents knowledge transfer and training strategy solutions for each: 1) Transfer learning to solve the lack of data problem for one target task, 2) Domain generalization that can test multiple target domains with small domain gap without additional training, 3) Continual learning: sequentially training multiple target domains with large domain gap while minimizing catastrophic forgetting. We performed transfer learning in the invisible mobile keyboard decoding task, domain generalization was studied in the face swapping task, and we applied continual learning to the unsupervised domain adaptation task for the first time in the world.