Self-Directed Lifelong Visual Learning
Abstract
Our work under Phase I of the DARPA Lifelong Learning Machines (L2M) Program has concentrated on solving lifelong learning problems in the domain of computer vision and machine learning. Our work has focused on solving the general task of Intelligent Search, which is the combined visual and navigation task of finding an object, described either by name or with a picture, in a new or unknown environment. There are many facets of this problem, including the reasoning of where an object is likely to be based on a visual assessment of the current scene, the ability to navigate quickly and efficiently through unknown environments, the inference of3D structure from images to aid navigation, and learning policies for efficient navigation. We report on progress for all four of these goals. In addition to these four areas of research being directly related to the larger goal of Intelligent Search, they are all central topics in Lifelong Learning. In each area, the agent is designed so that it continues to improve as it exposed to new distributions of objects, new types of environments, and new types of obstacles. Our results include developing agents that are able to learn models of context to more efficiently find unseen objects, agents that are able to more efficiently navigate in unknown environments, and an extensive and highly detailed 3D scene database with substantially more detail than previous data sets. We detail a large number of scientific publications that describe and support this work.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 2020
- Accession Number
- AD1115179
Entities
People
- Eric Learned-miller
Organizations
- University of Massachusetts Amherst