Abstract

To support real-time &sustainable machine intelligence that exploits the rapid growth in sensor deploymentsin urban spaces (e.g., video, audio) and on wearable devices (e.g., inertial, radar), there is a need to optimize the execution of machine learning (ML) pipelines on resource-constrained embedded devices. To this end, this talkshall describe the vision of collaborative machine intelligence, where the sensing and inferencing pipelines on individual wearable and IoT devices collaborate, in real-time, to overcome such resource limitations. First, I will describe work on tightly coordinated IoT+ wearable sensing, which allows the ultra-low power (even battery-less) capture of fine-grained human gestural activities in various environments (e.g., in offices and in gyms) by combining IoT sensors &wearable devices. Second, using a sample video surveillance application, I will describe how IoT-based collaborative machine inferencingcan provide dramatic reductions in energy and latency, as well as improvements in accuracy. To practically realize this vision, I shall finally argue why edge computing needs to evolve, from its current focus on pure local computation offloading to a ``Cognitive Edge” platform that enables such collaborative and trusted sense-making across heterogeneous pervasive devices.

Abstract

The coming decades may see the large scale deployment of networked cyber–physical systems to address global needs in areas such as energy, water, health care, and transportation. However, as recent events have shown, such systems are vulnerable to cyber attacks. We begin by revisiting classical linear systems theory, developed in more innocent times, from a security-conscious, even paranoid, viewpoint. Then we present a general technique, called "dynamic watermarking," for detecting any sort of malicious activity in networked systems of sensors and actuators. We present a field test on an automobile, experimental demonstration of this technique on an automobile on a test track, an experimental process control system, and a simulation study of defense against an attack on Automatic Gain Control (AGC) in a synthetic four area power system.

[Joint work with Bharadwaj Satchidanandan, Jaewon Kim, Woo Hyun Ko, Tong Huang, Gopal Kamath, Lantian Shangguan, Kenny Chour, Le Xie, and Swaminathan Gopalswamy].

Abstract

Semantically-rich visual information, generated by surveillance cameras or those on mobile devices, as well as by 3D sensors that provide depth perception (like LiDAR and stereo-cameras), is available aplenty at the network's edge. However, these sensors have limited communication bandwidth to the rest of the network, and sometimes limited on-board compute. This talk will cover experiences drawn from, and draw out common design patterns that arise in, designing systems that overcome these challenges to realize a range of interesting capabilities: extended vehicular vision, real-time visual map updates, cross-camera complex activity detection, and visual analytics in retail settings.

Abstract

In every form of digital store-and-forward communication, intermediate forwarding nodes are computers, with attendant memory and processing resources. For more than 30 years this has stimulated efforts to create a wide-area infrastructure that goes beyond simple forwarding to create a platform that makes more general and varied use of increasingly powerful and plentiful node resources. There have been analogous pressures toward active and networked storage, processor-in-memory and streaming processors. There is a widespread consensus that it should be possible to define and deploy a converged wide-area platform that combine these silos seamlessly and universally. However a great deal of investment in research prototypes has yet to produce a credible candidate architecture. Drawing on design analysis, historical examples, and case studies, this talk presents an argument for the hypothesis that in order to realize a distributed system with the kind of convergent generality and deployment scalability that might qualify as "future-defining," we must build it from a small set of simple, generic, and limited abstractions of the low level resources (processing, storage and network) of its constituent nodes. The common building block out of which these silos are constructed are storage or memory buffers/blocks and a set of primitive allocation and processing operations on them.