Abstract:

Today, videos captured and analyzed by computer vision models grow faster in volume than videos watched by human viewers. Despite that, traditional video delivery and processing systems are designed for human perception, whereas computer vision models (deep neural nets or DNNs) “perceive” video data differently—they value high inference accuracy and low inference delay. This discrepancy in performance objectives has far-reaching consequences. In this talk, I will introduce two new abstractions that enable video analytics systems to self-adapt to dynamic characteristics of input video streams and the idiosyncrasies of different DNN models. The key insight is that, while eliciting real-time feedback from human users is hard, DNN models can naturally provide rich and instantaneous feedback that could be harnessed to guide efficient system adaptation.

Abstract:

Emerging applications such as smart cities, autonomous vehicles, and mixed reality rely on embedded systems that are engaging with the physical environment through sensors. Building upon this connection, my vision is to advance Omnipresent Sensing by harnessing the wireless infrastructure in and around buildings and cities to act as a non-intrusive sensing platform. This is possible by innovating at the crossroads of two recent trends in mobile computing: (1) Wireless technologies such as Millimeter-wave (mmWave) and Massive MIMO systems can now support higher bandwidth for communication and improved resolution for RF sensing applications. (2) Advancements in CPUs and RF front-ends are making it easier to develop software-defined sensing and communication systems. This is affording edge devices the ability to do more advanced signal processing and machine learning.

In this talk, I will focus on how to design an RF-equivalent of optical retro-reflectors and use them as fiducial markers in autonomous vehicles, robotics, and mixed reality applications. I will then discuss how nuances from the environment itself can be leveraged to improve sensing quality in the context of human sensing, object tracking, and indoor localization. I will conclude this talk with a roadmap of combining radar-style RF sensors and wireless communication links for the wireless embedded systems of the future.

Abstract:

The mobile networks (4G, 5G, and beyond) facilitate "anywhere, anytime" Internet access for billions of users and trillions of smart "things." Despite this great success, however, the mobile networks remain a complex "black box." This design is vulnerable to various security threats and reliability/performance deficiencies. It also hurts the fundamental mutual trust between mobile users and operators.
This talk will show how end clients can help strive for a smart and secure mobile network. We will start with a data-driven approach to let smart clients understand the complex and "black-box" mobile network operations. This new capability empowers various IoT/mobile applications, and rewards the 4G/5G infrastructure with simplified yet verifiable designs and operations. We will demonstrate this with two examples: Enhancing the mutual trust between users and operators in the intelligent edge, and resolving the policy conflicts in the distributed mobility management. We last discuss some research directions toward smart and secure 6G and future mobile Internet.

Abstract:

Smart infrastructures are usually driven by intelligent context-aware services that often need strong ML (or DL) models working in the backend. However, building such sophisticated models often needs a significant amount of labeled training data. The most accepted and the conventional source of such data annotation considers human-in-the-loop. Nevertheless, this task is getting expensive and often highly tedious with multimodal data sources in place. Considering the broad use case of sensor-based human activity recognition, in this talk, we first discuss the possibility of choosing auxiliary modalities that can provide enough information to get the data annotated without human-in-the-loop. Subsequently, we also look into zero-shot approaches to fine-tune the coarse-grain annotations received from external annotators to capture more information regarding the confounded actions hidden within any complex activity of daily living.

Abstract:

We need affordable and innovative sensing to assist in tackling global-scale problems like sustainably feeding the next billion. RF backscatter systems such as RFID are well-known in the sensing community because of their low power consumption. This talk explores how we can enable robust indoor/outdoor RF-backscatter sensing by leveraging advances in low-power embedded systems. The talk will describe the challenges faced in low-power and sustainable sensor networks, and provides an overview of an outdoor radar backscatter sensing system that uses RF to measure soil moisture with accuracy comparable to state-of-the-art commercial sensors, but at a fraction of the cost. Also to be discussed is recent work on renewably powering outdoor sensor systems via energy harvested from non-traditional sources like microbes.

Abstract:

For long, mobile sensing and AI research have focused on creating accurate, robust, and multi-modal sensory models, making it intelligent to understand us and the world around us. However, with an unprecedented rise of on/near-body devices - smartphones, wearables, or IoT devices - it is common today to find ourselves surrounded by multiple sensory devices in daily lives. Such multi-device dynamics offer exciting opportunities to leverage multiple, diverse sensory signals to develop accurate and robust sensing models quantifying diverse and richer human context continuously and at scale. In this talk, I will explore system challenges and research efforts in building sensory AI systems to support multi-device environments. Specifically, I will cover the systems issue from three key perspectives: a) device form, b) device intelligence, and c) device system.

Abstract:

This talk will provide an overview of the COSMOS testbed (www.cosmos-lab.org), that is being deployed as part of the NSF Platforms for Advanced Wireless Research (PAWR) program, and the supported experiments in advanced wireless, millimeter-wave (mmWave), optical networking, and edge cloud. COSMOS (Cloud-Enhanced Open Software-Defined Mobile-Wireless Testbed for City-Scale Deployment) is being deployed in West Harlem (New York City) by Rutgers, Columbia, and NYU in partnership with NYC, CCNY, U. Arizona, IBM, and Silicon Harlem. It targets the technology “sweet spot” of ultra-high bandwidth and ultra-low latency, a capability that will enable a broad new class of applications including augmented/virtual reality and cloud-based autonomous vehicles. Realization of such high bandwidth/low latency wireless applications involves research not only on faster radio links, but also on the system as a whole including aspects such as spectrum use, networking, and edge computing.
We will present an overview of COSMOS’ key enabling technologies, which include millimeter-wave (mmWave) radios, software-defined radios, optical/SDN x-haul network, and edge cloud, followed by the deployment and outreach efforts. We will then describe extensive millimeter-wave channel measurements and experiments on open-access full-duplex wireless, as well as the development and integration of programmable mmWave radios in the COSMOS testbed. We will conclude by converged optical-wireless x-haul networking with edge cloud that we conducted in the COSMOS testbed.
The design, development, and deployment of the COSMOS testbed are joint work with the COSMOS team (www.cosmos-lab.org). The mmWave measurement results are based on joint work with Manav Kohli, Tianyi Dai, Angel Daniel Estigarribia, Gil Zussman (Columbia), Rodolfo Feick (UTFSM), Dmitry Chizhik, Jinfeng Du, and Reinaldo A. Valenzuela (Nokia Bell Labs). The full-duplex results are based on joint work with Mahmood Baraani Dastjerdi, Manav Kohli, Harish Krishnaswamy, and Gil Zussman (Columbia). The programmable mmWave radios are joint work with Xiaoxiong Gu, Arun Paidimarri, Sadhu Bodhisatwa, Alberto Valdes-Garcia (IBM Research), Gil Zussman (Columbia), and Ivan Seskar (Rutgers/WINLAB).