About SAVN

SAVN Teaser Figure
Fig. 1. Traditional navigation approaches freeze the model during inference (top row); this may result in difficulties generalizing to unseen environments. In this paper, we propose a meta-reinforcement learning approach for navigation, where the agent learns to adapt in a self-supervised manner (bottom row). In this example, the agent learns to adapt itself when it collides with an object once and acts correctly afterwards. In contrast, a standard solution (top row) may make multiple mistakes of the same kind when performing the task.

When humans learn a new task there is no clear distinction between training and inference. Even after we explicitly learn a task, we keep learning while we perform. In this project we study the problem of learning to learn without supervision in the context of visual navigation. Our agent learns how to adapt itself to environments by interacting with them. Accordingly, the agent may continue learning even during inference.

Model

SAVN Model
Fig. 2. Our network optimizes both a self-supervised interaction loss and navigation loss. The network outputs a policy given the egocentric image and word embedding of the target object class. During training, the interaction and navigation-gradients are back-propagated through the network, and the parameters of the self-supervised loss are updated at the end of each episode using navigation-gradients. At test time the parameters of the interaction loss remain fixed while the rest of the network is updated using interaction-gradients. Note that the green color in the figure represents the intermediate and final outputs.

Code & Data

Paper

Learning to Learn How to Learn: Self-Adaptive Visual Navigation using Meta-Learning

Learning to Learn How to Learn: Self-Adaptive Visual Navigation using Meta-Learning

Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi • CVPR • 2019

Video