2018/12/03 by Mitchell Wortsman, Wortsman, Mitchell, Kiana Ehsani +7 · 8 citations
Computer Science · #Advanced Vision and Imaging #Domain Adaptation and Few-Shot Learning #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.1812.00971
Learning is an inherently continuous phenomenon. When humans learn a new task\nthere is no explicit distinction between training and inference. As we learn a\ntask, we keep learning about it while performing the task. What we learn and\nhow we learn it varies during different stages of learning. Learning how to\nlearn and adapt is a key property that enables us to generalize effortlessly to\nnew settings. This is in contrast with conventional settings in machine\nlearning where a trained model is frozen during inference. In this paper we\nstudy the problem of learning to learn at both training and test time in the\ncontext of visual navigation. A fundamental challenge in navigation is\ngeneralization to unseen scenes. In this paper we propose a self-adaptive\nvisual navigation method (SAVN) which learns to adapt to new environments\nwithout any explicit supervision. Our solution is a meta-reinforcement learning\napproach where an agent learns a self-supervised interaction loss that\nencourages effective navigation. Our experiments, performed in the AI2-THOR\nframework, show major improvements in both success rate and SPL for visual\nnavigation in novel scenes. Our code and data are available at:\nhttps://github.com/allenai/savn .\n