A Study of Deep Cortical Processing Algorithms for Advancing Next-Generation Machine Learning Systems
Abstract
Statement of Work: This project will focus on key challenges in machine intelligence by developing a new set of algorithms and architectures that can effectively address next-generation problems. The overarching goal is to provide side-by-side comparisons of current state-of-the-art schemes and the enhanced machine learning algorithms and architectures on a set of standardized benchmarks. The broader goal of this work is to not only demonstrate the improvements to the current body of work in biologically-inspired machine learning but also pave the way for future research and development aimed at further advancing machine intelligence technologies. Objective: (a) Develop algorithms for mitigating catastrophic forgetting —that is, for enabling learning mechanisms that do not result in loss of prior knowledge. (b) Develop algorithms for real-time adaptation in a biologically-inspired manner which stands in contrast to lengthy offline learning that characterizes most existing machine learning schemes. (c) Investigate representations that capture temporal dependencies which span broad time scales – a key property when dealing with natural signals, including visual, auditory and biomedical. Approach: The investigators will identify publicly available benchmarks (and perhaps introduce new ones) that clearly underscore the three key machine learning objectives of this effort, and establish state-of-the performance profiles of existing hierarchical/cortical learning techniques. In particular, we will study real-time adaptation to support incremental learning and task transfer. Moreover, the team will conduct side-by-side comparisons that will clearly demonstrate the value of the developed enhancements in the context of both target tracking and activity recognition application domains. Detailed reports will be generated, highlighting the findings as well as suggesting future research directions to further advance machine intelligence technologies. Overall Merit and ONR Mission/Relevance: The project will support ONR s mission of building intelligent agents that can function in environments in which warfighters operate, that is, environments that are unstructured, open, complex and dynamically changing. In particular, the effort will contribute to technologies pertaining to autonomous agents that require learning complex concepts and tasks from examples, instructions, and demonstrations. The effort will also contribute to the area of image understanding, particularly in recognizing and tracking objects in challenging warfare settings, such as poor lighting conditions, noisy data feeds and partial observability.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512615
Entities
People
- Silvio Savarese
Organizations
- Office of Naval Research
- Stanford University
- United States Navy