Optimizing for Intelligent Embedded Devices
Abstract
A major technological trend of the past several years is the emergence of commercially available Internet-enabled embedded devices, which together are forming the Internet of Things. These devices include personal wearables such as smart wrist watches and healthcare devices, as well as larger devices such as low-altitude UAVs and Lidar-sensor enabled devices. With the continued support of industry and techniques from researchers, these devices can perform intelligent tasks such as learning unknown environments and targets of interest, 3D mapping surroundings, performing data analytics, and offering predictions, all enabled by computational models for sense making. In particular, the developments realized via these devices will deliver unprecedented technological capabilities to address the tasks described in NPS-BAA-15-004. By offering information and actionable predictions in real time, these devices will expand the capacity of human agents analyzing under-sampled and deceptive target systems, and allow them to perform in environments characterized by the high dimensional data and information produced by these devices. As such, these devices will both complement the existing and enable the future orchestrated intelligence infrastructure by empowering human-computer symbiotic models. However, in order to realize this potential, we need to address the following challenge: How can an embedded device accomplish these tasks with a small hardware and software footprint, in near real time, and with low power consumption? The answers to these challenges are difficult, as there are gaps in the mathematical models and methods in areas such as integrating data from multiple modalities, performing complicated computations using limited resources of embedded devices, and recovering important information from device data obtained via low-energy signal acquisition which may be subject to nonlinear distortion. New approaches are needed, as most of the big data approaches developed over the past decade, such as deep learning, are datacenter-based. There is a lack of theoretical and algorithmic work to link these efforts to embedded devices. In this effort, we propose new research directions, under a general sparse-coding based framework, to bring intelligence to these compute and energy limited devices. We develop novel methods of leveraging multiple data modalities to learn discriminative dictionaries, selecting client-specific dictionaries for performing common ISR tasks, and using nonlinear compressive sensing for energy-efficient data acquisition, all aimed at equipping embedded devices of limited resources with intelligence. The proposed work will be conducted by the PI with his postdocs and graduate students at Harvard. The research grant will be used solely to support basic open research in the proposed areas of work, with all results published in open literature and disseminated in research communities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2016
- Source ID
- N002441610018
Entities
People
- H.t. Kung
Organizations
- Harvard University
- United States Air Force