THIS IS A CONTINUATION OF N00014-13-1-0720 Multiscale Learning for Integrated Scene Understanding
Abstract
Short Work Statement:Develop a new class of structured learning algorithms that can learn complex representations, reason efficiently over them, incorporate varying amounts of supervision and domain knowledge, and induce multiple levels of features. These algorithms will scale to large, high-dimensional datasets such as those found in vision. The focus of these learning algorithms is scene understanding.Objective:The objective of this project is to develop algorithms for unprecedented levels of scene understanding throughdeveloping a new class of structured learning algorithms, capable of learning complex representations, reasoning efficiently over them, incorporating varying amounts of supervision and domain knowledge, and inducing multiple levels of features; and scaling these algorithms to very large, very high-dimensional datasets such as those found in vision.Approach:The key to effective structured learning is representations that are as expressive as possible while remaining tractable or nearly so. In turn, multiscale structure provides a foundation for such representations. Dimensions of scale include space and time as well as class and part hierarchies, both within a learning episode and across the learners history. The core of this project will be to develop a series of increasingly powerful multi-scale learning algorithms, capable of inducing the full spectrum from low-level features to high-level relational knowledge. To accomplish these, the Pis will develop methods for: (a) Learning tractable Markov logic networks; (b) Learning the structure of sum-product networks and their distributions over manifolds; ~ Approximate inference; (d) dynamic feature acquisition. The scale required to make these algorithms practical for scene understanding can only be achieved throughdistributed computation. To this end, the Pis will build on the GraphLab project, which allows distributed learning and inference on structures that can be represented as graphs, such as MLNs and SPNs. The Pis will address the following fundamental problems: (a) Discovering and changing graph structures; (b) Distributed computation without graphs; ~ Distributed structured prediction cascades: Using; (d) Persistence and fault tolerance.Scene understanding entails reasoning in terms of several interlacing components, but most work to date has focused on learning them individually. Considering them jointly is extremely complex, but the structured learning algorithms the Pis will develop in this project will make it possible. In particular, their integrated approach to scene understanding will include the following components: (a) Combining recognition and reconstruction; (b) Occlusion reasoning; ~ Taming intra-class variance; (d) Multiscale sparse coding; and Structured scene summarization.Overall Merits and ONR Relevance:This work is expected to substantially contribute to the methods for learning complex concepts including complex objects and their relationships, and scenes.This work is highly relevant to enabling Navy s Autonomy and Information Dominance. Automated methods for recognizing objects and scene are of critical importance to naval missions that include perception for autonomous agents and understanding surveillance imagery.Progress:This project is a collaboration of Pedro Domingos, Ali Farhadi, Dieter Fox and Carlos Guestrin. A main direction this year was extending sum-product networks (SPNs) and Tractable Markov logic (TML) to handle increasingly challenging, large-scale, real-world problems in vision and learning. Our paper on using SPN ideas for non-convex optimization and applying them to bundle adjustment and protein folding won the IJCAI-2015 Distinguished Paper Award. We developed, tested and published the first full-blown implementation of TML, including new algorithms for highly scalable inference and learning, and their application to building and exploiting a probabilistic knowledge base with millions of objects and billions of p
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141612697
Entities
People
- Pedro Domingos
Organizations
- Office of Naval Research
- United States Navy
- University of Washington