Upamanyu Madhow - US grants

The research studies the design of agile wireless networks that
accommodate time variations in the communication channels, the
information sources, and the network topology.
The research will lead to design principles that, in addition to
enabling more efficient use of the current cellular and PCS bands, will allow
exploitation of frequency bands in the 10-100 GHz range to provide high-speed
multimedia services for both indoor and outdoor applications. While the basic
cellular paradigm of wireless access to a high-speed communication and
computing backbone will be adhered to, nearly every other assumption
in existing second-generation and projected third-generation cellular
and PCS networks will be reexamined. Some of the significant differences
from current designs are as follows. A dense network of base stations
will provide connectivity despite the high path losses
and sharp shadowing at higher frequency bands. Cells with
well-defined boundaries may no longer exist, and mobile terminals will
see a rapidly varying network topology. A variety of traffic classes,
such as voice, data, and video, with diverse requirements regarding
delay, loss, quality of reproduction, and number of potential receivers
will be considered.

The integrated research effort is being conducted by five overlapping
research teams of University of Illinois faculty investigators and
their students, organized around the following interdisciplinary projects:
(1) Concept Systems, Modeling, and Performance Limits,
(2) Design Principles for Wireless Packet Networks,
(3) Design for Time-Varying Channels,
(4) Jointly Optimized Source Coding, Channel Coding, and Estimation, and
(5) VLSI Algorithms, Architectures, and Bounds.
The investigators are applying their expertise in digital signal
processing, communication systems, networking, circuit design and control.
A diverse array of methods will be applied, including modeling
and simulation, application of adaptive and universal algorithm
methodologies, exploitation of antenna arrays and polarization for
diversity and beamforming, algebraic coding techniques, linear and nonlinear
optimization techniques, information theory, and laboratory
implementation and measurement.

EIA-0080134
Singh, Ambuj
University of California - Santa Barbara

CISE Research Infrastructure: Digital Campus: Scalable Information Services on a Campus-Wide Wireless Network

Researchers at the University of California at Santa Barbara will implement a wireless-networked, distributed heterogeneous environment on campus and use it to conduct research in databases, networking, distributed systems, and multimedia. The PIs will focus on large-scale systems in which data is the critical resource and system services are based on various data manipulation functions including data collection, movement/delivery, aggregation/processing, and presentation. A significant part of the research will be conducted using a digital classroom, a remote classroom, and individual and team kiosks. Services such as lecture on demand, virtual offices, and remote learning will be provided using this infrastructure. Specific research issues that will be investigated include content-based access, personalized views, multi-dimensional indexing, smart end-to-end applications, joint source-network coding, scalable storage, reliable network service, information summarization, distributed collaboration, multimedia annotation, and interactivity.

This project addresses some fundamental aspects of the theory and practice of wireless networking. An integrated approach, combiningphysical layer innovations with new protocols for medium access control and scheduling, while accounting for application requirements and transport protocol dynamics, is employed for solving the research problems that are identified. Two major research thrusts are considered. In the first research thrust, the concept of ``pseudocellular'' wireless networks, which combine the best features of cellular and ad hoc networks, is considered as a paradigm for plug-and-play fourth generation (4G) wireless networks. Such a flexible architecture is clearly critical for quick set-up of wireless networks in emergency situations, in which stationary, or perhaps even mobile, base stations are deployed at convenient (but not optimized)
sites to serve both slow-moving and fast-moving users. However, it is also a key ingredient of our vision of achieving a quantum jump in wireless link speeds, by going beyond the current cellular frequency bands of 1-2 GHz to the large bandwidths available in frequency bands in the 10s of GHz. The path loss in such bands is high, forcing the use of a dense network of base stations on the one hand, and enabling more aggressive frequency reuse on the other. The focus of the research is to support a mix of user mobilities, and a mix of real-time and non real-time applications, over a packetized pseudocellular infrastructure. This setting differs from conventional cellular networks, in that the cell sizes are small, and cells may have substantial overlap. It differs from wireless Local Area Networks (WLANs), in that it allows for rapidly mobile users despite the small cell sizes. Instead of a conventional hierarchical structure (i.e., large cells for fast-moving users, overlaid on small cells for slow-moving users) to deal with a range of mobility, a mobile-centric approach, which combines handoffs and reservation-based medium access control, is considered to allow for flexible deployment. A novel idea to be investigated is the support of priorities on the reservation channel, so as to allow, for example, highly mobile users with real-time calls in progress to rapidly reserve resources when entering a new pseudocell, thus implicitly achieving a handoff. Another important issue is transceiver optimization of the reservation channel, which requires solution of new problems in multiuser communications.

The second research thrust is motivated by the well-known observation that Internet traffic has a heavy-tailed distribution, which typically calls for more conservative resource provisioning than for traditional Markovian traffic models. Since overprovisioning is unattractive in resource-constrained wireless environment, the approach considered is toemploy a new Quality of Service (QoS) framework that allows foraggressive resource utilization, by serving the bulk of the
transactions (which are short) rapidly, and penalizing the small fraction of long transactions that contribute to the heavy tails. Scheduling disciplines that achieve this goal are very different from
popular round robin or fair queueing schedulers, and were considered in the queueing theory literature more than three decades ago. The implication of these results for heavy-tailed Internet
traffic is explored for the first time (to the best of our knowledge) in this project. The scheduling strategy is extended to a shared wireless channel, where fairness is traded off against system
efficiency, with the latter dictating that users seeing the best channels are the ones that should get link access. The tradeoff is expected to be biased towards efficiency in order to support
heavy-tailed traffic effectively. The interaction between scheduling and the dynamics of TCP connections (TCP is the Internet data transport protocol on top of which most transactions run) is explored, keeping in mind that a TCP connection that is starved of network resources can get locked out of the network due to repeated timeouts and rate cutbacks. Finally, the dependence of scheduling on mobility is explored, with the concept of assigning priority to highly mobile
users (who have a smaller chance of getting access to the link during their sojourn in a given pseudocell), while keeping overall QoS and fairness in mind. The scheduling methods we develop place a high importance on overall system efficiency, and are therefore well-suited to flat rate pricing, which is arguably an effective mechanism of promoting usage growth in wireless data networks.

Wireless sensor networks have a large array of potential applications,
including environmental monitoring, home and factory automation, and
homeland security. This research addresses efficient data collection
from sensor networks, which is one of the fundamental bottlenecks to
large-scale deployment of low-cost sensor nodes. Two key difficulties
are scale (how to manage very large numbers of relatively
unsophisticated nodes) and energy (communication is energy-intensive,
but sensor nodes must operate for months or years on battery, or by
scavenging energy from the environment). This research has two
complementary research thrusts, both investigating novel techniques
for data collection from sensor networks, exploiting the natural
distribution of sensor nodes in space. The first, termed virtual
radar, enables deployment of extremely simple sensor nodes which do
not need networking or position location functionality, by moving the
complexity to a sophisticated data collection node. The second
approach requires more sophisticated nodes, and increases energy
efficiency and range by obtaining the functionality of multi-antenna
transmission by intelligent coordination between neighboring
(single-antenna) sensor nodes.

In the first thrust, termed virtual radar, sensor nodes with activity
to report electronically reflect a beacon transmitted by a
"driveby" or "flyby" collector node, thus
creating a radar-like geometry. The collector employs sophisticated
synthetic aperture radar like signal processing techniques to obtain
an "image" of the activity in the sensor field.
Tracking of events moving across the sensor field, and exploiting
multiple antennas at the collector, are investigated. The second
thrust, termed distributed transmission, involves investigation of
beamforming, diversity and power gains obtained by coordinating
between neighboring nodes, as a function of the level of
synchronization achieved among the nodes. Methods of achieving such
synchronization are also considered.

Large scale wireless sensor networks with tens of thousands of nodes have a host of potential applications, including homeland security, environmental monitoring and planetary exploration. However, conventional multihop wireless networks do not scale to such large numbers of nodes, and node localization is difficult for random deployment of ultra low-cost sensors. This research provides a proof-of-concept for Imaging Sensor Nets, which address the problems of scale and localization simultaneously, employing ideas analogous to GPS, RFID and CDMA. A "smart" collector node sends an RF beacon, which is electronically reflected and data-modulated by "dumb" sensors illuminated by the beacon. The collector employs baseband and radar/imaging algorithms to process these reflected signals in order to simultaneously estimate the locations and data of the sensors. Millimeter wave carrier frequencies are employed in order to enhance resolution for localization of sensor nodes. The research activities
include the challenging task of low-cost CMOS IC implementation of sensor hardware at millimeter wave frequencies (which are an order of magnitude higher than commercial RF communication systems), hardware brassboarding of the collector transceiver, design and software implementation of innovative baseband signal processing algorithms for timing acquisition and demodulation at the collector, and design and software implementation of imaging algorithms at the collector. In addition to the dissemination of results via the standard avenues of publications and the Internet, both indoor and outdoor demos of the technology will be widely publicized to industry and funding agencies in order to establish a clear path to technology transfer.

ECS-0636594
Chik Yue, Carnegie Mellon University
ECS-0636621
Upamanyu Madhow, University of Santa Barbara

Our objective is to develop the system architecture, signal processing algorithms and integrated circuit techniques for a robust, quick set-up, point-to-point wireless link which achieves speeds of 10-40 Gbps over a range of several kilometers, using millimeter (mm) wave spectrum. Since these speeds are comparable to those of optical fiber, the outcome of this project enables a fail-safe hybrid communication backbone infrastructure, which can be deployed or restored rapidly in the events of disaster and emergency. The system employs a novel hierarchical architecture which meshes beamforming (to provide link margins sufficient to overcome the limitations of mm-wave propagation in harsh weather) and spatial multiplexing (to provide large spectral efficiency, of the order of tens of bits per second per Hertz, required to realize optical link speeds using channel bandwidths of only several GHz). Beamforming gains are obtained by electronically steerable monolithic arrays. Each such array is a subarray in a larger array, forming a spatially multiplexed virtual multiple-input, multiple-output (MIMO) system: the transmit subarrays send separate data streams, which are separated out at the receiver using spatial interference suppression techniques. Key elements of this mm-wave MIMO system are CMOS IC design for monolithic steerable sub-arrays, signal processing/hardware co-design to obtain algorithms implementable at such high speeds, and hybrid analog/digital processing to enable low-power operation. Substantial effort will go into establishing a cell-based, reusable design/modeling framework to enable CMOS mm-wave VLSI design. The new findings will be incorporated into undergraduate and graduate classes through small design projects.

Intellectual Merit: This is an inherently interdisciplinary project whose success depends critically on intense interaction between the three PI's on this project, whose combined expertise spans CMOS IC design for communication applications (Yue), millimeter wave device and IC design (Rodwell) and signal processing for communication (Madhow). The proposed system is based on innovations at every level, including system concept, signal processing algorithms, and circuit design and packaging. Millimeter-wave MIMO provides spatial multiplexing in line of sight environments, and is therefore a completely new concept relative to MIMO at lower frequencies, which provides spatial multiplexing only in rich scattering environments. The electronically steerable sub-arrays are based on a unique row-column architecture amenable to monolithic realization. The innovation in the signal processing consists of drastic simplifications, including a hierarchical decomposition co-designed with the hardware. Circuit design at mm-wave frequencies push the limits of mixed signal design in low-cost CMOS processes, and our cell-based design framework has the potential of providing a systematic approach to such design. The baseband processing employs novel hybrid analog/digital processing techniques, in order to minimize the performance requirements on high-speed, high-cost, high-power analog-to-digital converters.

Broader Impact: Millimeter-wave MIMO provides the first feasible approach to bridging the capacity gap between wireless and optical systems, which has applications ranging from homeland security (e.g., disaster recovery) to last mile connectivity for enterprise and residential settings. An additional breakthrough is in terms of the ease of deployment of LOS outdoor links, which becomes a simple operation of roughly pointing the transmitter and receiver at each other, rather than precisely aligning the transmit and receive antennas as done in current practice. In addition, the breakthroughs in mm-wave CMOS circuit design and packaging required by this demanding application have the potential for impact well beyond the specific system considered here, and will open up a host of opportunities for harnessing mm-wave spectrum at reasonable cost. The PIs all have strong records of technology transfer, and intend to leverage their strong contacts with the communications industry to push for technology transfer by widely disseminating the results of this work not only through publications, but also using hardware demonstrations easily accessible to visitors. The proposed research will have a significant impact on the undergraduate and graduate curriculum at the PIs' institutions in terms of driving innovations and updates in a number of courses in circuit design and communication systems. Well-established outreach mechanisms in the nanotech area will be used to involve women and minorities, including high school students, in this effort. Due to the inherently interdisciplinary nature of this project, the students involved will receive a broad education cutting across several areas of Electrical and Computer Engineering.

The economies of scale of cellular and WiFi networks are enabled by low-cost integrated circuit implementations of sophisticated digital signal processing (DSP) algorithms in wireless communication transceivers. An implicit assumption in this approach is that analog received signals can be converted to a reasonably faithful digital representation, an assumption that breaks down as link speeds increase to the point that high-precision analog-to-digital conversion (ADC) becomes too costly and power-hungry. This project involves the design of wireless networks in the latter regime: the goal is to design low-cost links operating at multiGigabit speeds (i.e., more than an order of magnitude faster than WiFi), exploiting large swaths of unlicensed spectrum in the 3-10 GHz band and the 60 GHz band. The research rethinks communication transceiver design, with the starting assumption that high-speed ADCs are ``sloppy.'' The research involves obtaining fundamental performance benchmarks using information theory, and devising DSP algorithms that achieve these performance benchmarks. The ultimate objective is to enable a quantum leap in the speed of wireless networks for the home and enterprise, while preserving the economies of scale associated with low-cost silicon implementations.

While conventional systems use 6-12 bits of ADC precision, this research considers the design of communication systems for low-resolution (1-4 bits) ADC, including Shannon theoretic benchmarks and algorithms for synchronization and equalization. Since high-speed digital-to-analog conversion is easier than ADC, precoding strategies which move complexity to the transmitter are investigated. The use of time-interleaved ADCs to attain higher precision, and hence higher dynamic range, is considered for both singlecarrier and multicarrier systems. The approach is to design receiver algorithms that jointly address mismatch between the component ADCs and the channel dispersion.

The large amount of unlicensed and semi-unlicensed bandwidth available for millimeter (mm) wave communication enable multi-Gigabit wireless networking that can potentially transform the telecommunications landscape.

Intellectual Merit: This research investigates the use of the unlicensed 60 GHz ``oxygen absorption'' band for providing a quickly deployable broadband infrastructure based on multi-Gigabit outdoor mesh networking. Millimeter wave links are inherently directional: the directionality is required to overcome the increased path loss at higher frequencies, and is feasible for nodes with compact form factors using antenna arrays realized as patterns of metal on circuit board. This project addresses the cross-layer design of mesh networks with such highly directional links, in which implicit coordination using carrier sense mechanisms cannot be relied on, and there is no omni-directional mode for explicit coordination. In addition, the research will investigate new design principles for directional medium access control, with the challenge being to coordinate nodes despite the deafness induced by directionality, while taking advantage of the drastically reduced spatial interference. The project will also study methods for network discovery and topology updates, the interactions between scheduling and routing; and the impact of oxygen absorption on network capacity and protocol design/performance.

Broader Impact: The principal investigators will develop publicly available mm wave network simulation tool, intended to engage a larger research community in this emerging field. The investigators will also explore other mechanisms for broader impact including technology transfer, undergraduate research, and curriculum updates featuring mm wave communication.

Over the past two decades, advances in multi-antenna wireless systems have opened up the possibility of tremendous increases in wireless communication throughputs. The practical realization of these increases in real-world wireless networks is constrained by size and cost considerations that limit the number of antennas, especially for mobile devices. This research aims to achieve the gains from multi-antenna techniques in a distributed fashion by having groups of wireless transceivers pool together and cooperatively act like a virtual antenna array. The concept of virtual arrays has broad transformative potential for wireless networks by extending the numerous benefits of multi-antenna systems to networks of single-antenna devices. We will prototype and demonstrate virtual array techniques experimentally on a software-defined radio platform, and we plan to share our implementations as reusable building blocks to stimulate technology transition and to promote interactions between the academic, open-source software community and radio hobbyist communities. The elements of a virtual array have an unknown, typically time-varying, geometry and are driven by independent oscillators, each with stochastic drift; the main technical challenge is in maintaining distributed coherence in the array in the presence of these effects. While this is a very challenging problem, recent results have demonstrated the feasibility of virtual arrays.

This research will establish a solid theoretical foundation for distributed coherence and chart a clear path to technology transfer by applying the theory and techniques to the cross-layer design of concept systems based on virtual antenna arrays. Specifically we will develop a state space framework for tracking and prediction of oscillator dynamics and mobility, and scalable architectures for distributed transmission and reception appropriate for large distributed arrays of low-cost single-antenna devices. We will also identify fundamental tradeoffs and scaling laws for virtual arrays. We will apply and integrate the theory and techniques to two concept systems of great societal significance: Distributed base station provides multi-antenna capabilities even for low carrier frequencies where standard antenna arrays would be too bulky (e.g., white space frequencies). Distributed 911, enables a cluster of nodes to communicate with a possibly moving distant rescue vehicle which would be out of range for any one of the nodes. Our goal is to perform design and performance evaluation in sufficient detail to clear conceptual hurdles for implementation.

Smart phones and tablets enable consumers to enjoy rich audio and video content on the go, but the proliferation of such increasingly sophisticated mobile devices has created a capacity crisis for mobile operators. It is estimated that supporting rich media content for a rapidly increasing fraction of mobile users requires a 1000-fold increase in cellular network capacity, which current cellular bands simply cannot support. The research pursued under this grant explores an alternative, and potentially transformational, approach to cellular data, using unlicensed spectrum in the 60 GHz band, where the available bandwidth is orders of magnitude higher than those used in existing systems, at the level of multiple Gigabits per second throughput on the downlink to the mobile devices.

Base stations for the envisioned network will be deployed opportunistically (e.g., on lampposts and rooftops). Due to the small carrier wavelength, many antenna arrays with a very large number (e.g., 1000) of elements can be built into base stations which are no larger than a typical WiFi access point. Such antenna arrays can be used to direct pencil beams at mobile users, with peak data rates of multiples of Gigabits per second (order of magnitude higher than the highest WiFi data rates available today). However, the small carrier wavelength also implies that the radio waves are easily blocked by obstacles such as buildings, walls, and humans, including the body of the person carrying the mobile device. In order to handle such rapid changes in the propagation environment, novel techniques are developed for multiple base stations to coordinate, such that they can adapt their beams to maintain connectivity with a given mobile device, and can ensure that the data destined for the mobile follows it around. A novel asymmetric network architecture is employed, with low-bandwidth 60 GHz beaconing and multi-Gbps data on the downlink, and LTE feedback and lower-speed data on the uplink. The base stations employ compressive signal processing for rapid channel estimation and beam adaptation, based on the feedback from the mobiles. Distributed base station coordination mechanisms are developed for seamlessly switching base stations or paths. The architecture minimizes complexity and power consumption in the mobile device: the device's 60 GHz radio only needs to receive, and the device is oblivious of handoffs.

The mobile broadband capacity crisis is the greatest challenge facing cellular providers today, hence the success of this project can impact a multi-billion dollar industry. In order to maximize the potential for impact, the results and models will be widely disseminated to both industry and academia. The investigators plan significant efforts for recruitment and mentoring of female undergraduate and graduate students, organized around the concept of a caring community.

Wireless communication technologies such as cellular and WiFi are indispensable for modern society. However, existing wireless networks are under severe stress due to the explosive demand caused by smart mobile devices capable of creating and consuming large amounts of multimedia content (especially images and video). Meeting these demands is estimated to require 1000-fold increases in wireless network capacity, which cannot be obtained by incremental advances using existing spectrum. A promising approach for delivering the required revolutionary advances in wireless by employ the so-called 'millimeter (mm) wave' band, which has huge amounts of available spectrum (e.g., 7 GHz in the unlicensed 60 GHz band alone). The wavelength in these bands is an order of magnitude smaller than that in today's wireless networks, drastically changing the physical and propagation characteristics: for example, mm waves are easily blocked by obstacles such as human bodies, but steerable antenna arrays with a very large number of elements (up to 1000) can fit in compact form factors, enabling us to potentially steer around obstacles using bounces from reflectors. As a consequence, realizing the potential for mm wave communication requires a comprehensive reexamination of existing wireless design principles, using an interdisciplinary approach that goes all the way from antenna design to network protocols. The goal of this project is to take such an approach for establishing fundamental principles for design of next generation mm wave communication networks, with a research agenda combining cross-layer modeling, design, and performance evaluation, firmly grounded in experiment. A key technical issue is to how to efficiently adapt electronically steerable arrays with a large number of elements, and to integrate them into network protocols.

The research is driven by the following cutting edge system concepts: (a) Cellular 1000X, aimed at relieving the cellular capacity bottleneck via 60 GHz cellular links delivering Gbps data rates to the mobile, together with a seamless extension to indoor networks; (b) 'Wireless fiber' backhaul at 140 GHz for enabling Cellular 1000X, based on easy to deploy outdoor wireless mesh networks with link speeds approaching 40-100 Gbps; (c) 40 Gbps indoor 60 GHz links, aimed at going beyond nascent industry efforts such as NG60 that aim to upgrade link speeds in the recently developed IEEE 802.11ad wireless local area network standard. The goal of this project is to design a system that will achieve the stated objectives, and prototype an advanced proof-of-concept that will help pave the way for eventual technology transfer leveraging the close ties of the project team to industry. A 60 GHz experimental platform developed to support the research will be made available to the research community, to stimulate a broader academic effort in this area.

Due to the small carrier wavelengths, beamforming at both ends is critical to make the link budget work, but it is essential to make the beams electronically steerable to steer around obstacles (which ``look bigger at smaller wavelengths''), and to allow automatic network configuration. Cross-layer frameworks for resilient pencil beam networking for both Cellular 1000X and indoor WLANs will be developed and demonstrated. These will incorporate compressive array adaptation techniques, a core innovation to be demonstrated in this project. Compressive adaptation enables 3D beamforming for robust link budgets, steering around blockage, and spatial reuse, and enables scaling of both the number of antenna elements and the nodes in the network, unlike existing scan-based IEEE 802.11ad medium access control (MAC) techniques. System concepts to be designed and tested include (a) `Picocloud' network architectures that employ tight coordination between base stations and APs (for outdoor and indoor environments, respectively) to provide seamless connectivity in the face of blockage; (b) Integration of beamforming with spatial multiplexing in LoS or near-LoS environments, demonstrating the scaling of available degrees of freedom with carrier frequency through prototypes at 60 GHz and 140 GHz.

A reconfigurable phased array at 60 GHz will be developed and integrated with the NSF/CRI-funded WiMi software defined radio platform, in order to enable the preceding system-level explorations (while beamsteering ICs developed by industry have been incorporated into products, external control of the beamsteering coefficients is not available). In addition, a hardware testbed for LoS spatial multiplexing at 140 GHz will be developed to demonstrate the potential for 'wireless fiber' backhaul links beyond 100 GHz.

Machine learning has made tremendous advances in the past decade, and is rapidly becoming embedded in our daily lives. We experience its power directly when we interact with voice assistants and automated translation engines, which are improving rapidly every year. Machine learning tools also enable many of the functionalities underlying search engines, e-commerce sites and social media. Thus, machine learning has become an essential component of cyberspace and our interactions with it, and is now poised to enter our physical space, for example, as a core component of perception for autonomous vehicles and drones. Much of the recent progress in machine learning has been in the area of multilayer, or deep, neural networks, which can be trained to learn complex relationships by leveraging the availability of large amounts of data and massive computing power. However, before we rely on such capabilities for safety-critical applications such as vehicular autonomy, we must ensure the robustness and security of deep networks. Recent research shows, for example, that deep networks can be induced to make errors (e.g., to misclassify images) by an adversary by adding tiny perturbations which would be imperceptible to humans. This project develops a systematic framework for defending against such adversarial perturbations, blending classical model-based techniques with the modern data-driven approach that characterizes machine learning practice today. The project will be validated through two key applications of deep learning: image classification and speech recognition.

When the vulnerability of deep networks to adversarial perturbations was discovered a few years back, it was initially conjectured that this vulnerability is due to the complex and nonlinear nature of the neural networks. However, there is now general agreement that this vulnerability is actually due to the excessive linearity of deep networks. Motivated by this observation, this project aims to develop a systematic approach to study adversarial machine learning by utilizing the sparsity inherent in natural data for defense, and locally linear models of the network for attack. The proposed approach is based on exploiting signal sparsity to develop provably efficient defense mechanisms. In particular, the project first investigates a sparsifying frontend, designed to preserve desired input information while attenuating perturbations before they enter the neural network. This then leads to a defense mechanism based on sparsifying the neural network, with the goal of mitigating the impact of an adversarial perturbation as it flows up the network. The methodology brings together ideas from sparse signal processing, optimization, and machine learning, and aims to bridge the gap between systematic theoretical understanding and machine learning practice. The proposal has an extensive evaluation plan that focuses on two important real-world applications of adversarial machine learning: image classification and speech recognition.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Advances in low-cost low-power silicon radio frequency (RF) integrated circuits (ICs) in the last two decades have opened up the commercial applications for millimeter wave (mmWave) frequencies which are an order of magnitude beyond those used in WiFi and cellular today. Large-scale deployment of mmWave communication networks, such as NextG cellular infrastructure outdoors and NextG WiFi infrastructure indoors, implies that these resources can be leveraged for RF imaging at scales that are not otherwise possible. The project develops foundational algorithms, architectures and protocols for such Joint Communication and Imaging (JCAI) systems. Each sensor in such a system provides 4D measurements (range, Doppler, azimuth angle and elevation angle) whose resolution improves by going to higher frequencies. The project establishes US leadership in a critical technology by developing large-scale RF imaging using frequencies beyond 100 GHz. Outdoor applications include pedestrian and vehicular tracking for global situational awareness supporting vehicular autonomy, and addressing security challenges such as timely detection of illegal drones or unauthorized personnel. In indoor settings, the technology enables fine-grained inference/prediction of human actions for eldercare and smart home applications. RF imaging technologies are especially useful in low-light or high-smoke/fog conditions when visible light or infrared technologies are not effective.<br/><br/>The project develops and demonstrates a framework for JCAI at mmWave frequencies. A core aspect of the technical plan is to drastically improve resolution by synthesizing large apertures (Thrust 1). This employs a combination of novel approaches to single sensor design which utilize large antenna arrays developed for communication, and networked collaboration between multiple sensors. A complementary aspect (Thrust 2) is the strategic utilization of unmanned vehicles to image difficult-to-reach areas, utilizing the fixed infrastructure to reduce the robot payload. In Thrust 3, hardware at 140 GHz previously developed by the PIs for communication will be adapted to support demonstration of networked RF imaging at 100+ GHz. Thrust 4 develops a control plane for networked imaging, including a resource management framework based on imaging demand and imaging capacity, and protocols supporting collaborative imaging. The concepts and methods to be developed have potential impact in a vast array of applications, including vehicular autonomy and road safety, manufacturing automation, indoor and outdoor security, eldercare, and healthcare. The PIs will work closely with industry partners, building on their strong track record in transitioning mmWave research, and plan to incorporate this research into the undergraduate curriculum through courses, capstone projects, and REU projects.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Deep neural networks (DNNs) are attaining great success in an increasing array of applications, yet there remain persistent concerns regarding their lack of interpretability and robustness. The standard approach to training DNNs is to optimize an end-to-end cost function based on variants of gradient descent. This simple approach is flexible, allowing innovation in architectures and applications, and scaling to neural networks with a large number of parameters, given enough data and computational power. Such end-to-end, or top-down training, however, does not provide control over, or understanding of, the features being extracted by the layers of the neural networks. The vulnerability of DNNs to adversarial attacks, for example, is a symptom of this phenomenon. The proposed research seeks to address these drawbacks using ideas from communication theory and neuroscience: the goal is to actively shape the features being extracted by individual layers of the neural network, in addition to training the overall network to attain an end-to-end goal. This research will contribute to curricular enhancements in signal processing and machine learning explored via courses, REU projects and senior capstone projects.

The proposed technical approach leverages the existing computational infrastructure for training, while imposing layer-by-layer constraints aimed at producing sparse, strong activations. Drawing on ideas from communication theory, the goal is to learn “matched filters” which enhance the “signal-to-noise ratio (SNR)” at neuron outputs at each layer. One may show that this approach is consistent with Hebbian and anti-Hebbian (HAH) learning as posited in neuroscience, in which neurons that are strongly activated for an input are promoted, with less active neurons being demoted. This work posits enhanced robustness via such an SNR-maximizing strategy, together with additional nonlinear transformations such as divisive normalization borrowed from neuroscience. While preliminary visualizations indicate more interpretable neurons, there is reason to expect sparse, strong activations to be more amenable to quantitative interpretation via statistical and information-theoretic analysis. The goal of the proposed research is two-fold: to gain theoretical insight into HAH-based learning via toy models, and to demonstrate practical gains in robustness and interpretability relative to state of the art DNNs. Experimental evaluations will initially be conducted on image datasets which provide standard performance benchmarks.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.