Massive multi-user multiple-input multiple-output (MU-MIMO) wireless systems operating at millimeter-wave (mmWave) frequencies enable simultaneous wideband data transmission to a large number of users. In order to reduce the complexity of MU precoding in all-digital basestation architectures, we propose a two-stage precoding architecture that first performs precoding using a sparse matrix in the beamspace domain, followed by an inverse fast Fourier transform that converts the result to the antenna domain. The sparse precoding matrix requires a small number of multipliers and enables regular hardware architectures, which allows the design of hardware-efficient all-digital precoders. Simulation results demonstrate that our methods approach the error-rate of conventional Wiener filter precoding with more than 2x reduced complexity.
Millimeter-wave (mmWave) massive multi-user multiple-input multiple-output (MU-MIMO) technology promises unprecedentedly high data rates for next-generation wireless systems. To be practically viable, mmWave massive MU-MIMO basestations (BS) must (i) rely on low-resolution data-conversion and (ii) be robust to jammer interference. This paper considers the problem of mitigating the impact of a permanently transmitting jammer during uplink transmission to a BS equipped with low-resolution analog-to-digital converters (ADCs). To this end, we propose SNIPS, short for Soft-Nulling of Interferers with Partitions in Space. SNIPS combines beam-slicing---a localized, analog spatial transform that focuses the jammer energy onto a subset of all ADCs---together with a soft-nulling data detector that exploits knowledge of which ADCs are contaminated by jammer interference. Our numerical results show that SNIPS is able to successfully serve 65% of the user equipments (UEs) for scenarios in which a conventional antenna-domain soft-nulling data detector is only able to serve 2% of the UEs.
All-digital millimeter-wave (mmWave) massive multi-user multiple-input multiple-output (MU-MIMO) receivers enable extreme data rates but require high power consumption. In order to reduce power consumption, this paper presents the first resolution-adaptive all-digital receiver ASIC that is able to adjust the resolution of the data-converters and baseband-processing engine to the instantaneous communication scenario. The scalable 32-antenna, 65 nm CMOS receiver occupies a total area of 8 mm$^\text{2}$ and integrates analog-to-digital converters (ADCs) with programmable gain and resolution, beamspace channel estimation, and a resolution-adaptive processing-in-memory spatial equalizer. With 6-bit ADC samples and a 4-bit spatial equalizer, our ASIC achieves a throughput of 9.98 Gb/s while being at least 2x more energy-efficient than state-of-the-art designs.
All-digital basestation (BS) architectures for millimeter-wave (mmWave) massive multi-user multiple-input multiple-output (MU-MIMO), which equip each radio-frequency chain with dedicated data converters, have advantages in spectral efficiency, flexibility, and baseband-processing simplicity over hybrid analog-digital solutions. For all-digital architectures to be competitive with hybrid solutions in terms of power consumption, novel signal-processing methods and baseband architectures are necessary. In this paper, we demonstrate that adapting the resolution of the analog-to-digital converters (ADCs) and spatial equalizer of an all-digital system to the communication scenario (e.g., the number of users, modulation scheme, and propagation conditions) enables orders-of-magnitude power savings for realistic mmWave channels. For example, for a 256-BS-antenna 16-user system supporting 1 GHz bandwidth, a traditional baseline architecture designed for a 64-user worst-case scenario would consume 23 W in 28 nm CMOS for the ADC array and the spatial equalizer, whereas a resolution-adaptive architecture is able to reduce the power consumption by 6.7x.
Wireless communication systems that rely on orthogonal frequency-division multiplexing (OFDM) suffer from a high peak-to-average (power) ratio (PAR), which necessitates power-inefficient radio-frequency (RF) chains to avoid an increase in error-vector magnitude (EVM) and out-of-band (OOB) emissions. The situation is further aggravated in massive multiuser (MU) multiple-input multiple-output (MIMO) systems that would require hundreds of linear RF chains. In this paper, we present a novel approach to joint precoding and PAR reduction that builds upon a novel $\ell^p\!-\!\ell^q$-norm formulation, which is able to find minimum PAR solutions while suppressing MU interference. We provide a theoretical underpinning of our approach and provide simulation results for a massive MU-MIMO-OFDM system that demonstrate significant reductions in PAR at low complexity, without causing an increase in EVM or OOB emissions.
Beamspace processing is an emerging technique to reduce baseband complexity in massive multiuser (MU) multiple-input multiple-output (MIMO) communication systems operating at millimeter-wave (mmWave) and terahertz frequencies. The high directionality of wave propagation at such high frequencies ensures that only a small number of transmission paths exist between user equipments and basestation (BS). In order to resolve the sparse nature of wave propagation, beamspace processing traditionally computes a spatial discrete Fourier transform (DFT) across a uniform linear antenna array at the BS where each DFT output is associated with a specific beam. In this paper, we study optimality conditions of the DFT for sparsity-based beamspace processing with idealistic mmWave channel models and realistic channels. To this end, we propose two algorithms that learn unitary beamspace transforms using an $\ell^4$-norm-based sparsity measure, and we investigate their optimality theoretically and via simulations.
We introduce a framework for linear precoder design over a massive multiple-input multiple-output downlink system and in presence of nonlinear power amplifiers (PAs). By studying the spatial characteristics of the distortion, we demonstrate that conventional linear precoding techniques steer nonlinear distortions in the direction of the users. We show that, by taking into account PA nonlinearity characteristics, one can design linear precoders that reduce, and in single-user scenarios, even remove completely the distortion transmitted in the direction of the users. This, however, is achieved at the price of a considerably reduced array gain. To address this issue, we present precoder optimization algorithms which simultaneously take into account the effects of array gain, distortion, multiuser interference, and receiver noise. Specifically, we derive an expression for the achievable sum rate and propose an iterative algorithm that attempts to find the precoding matrix maximizing this expression. Moreover, using a model for PA power consumption, we propose an algorithm that attempts to find the precoding matrix minimizing the consumed power for a given minimum achievable sum rate. Our numerical results demonstrate that the proposed distortion-aware precoding techniques yield considerable improvements in terms of spectral and energy efficiency compared to conventional linear precoding techniques.
Low-resolution neural networks represent both weights and activations with few bits, drastically reducing the multiplication complexity. Nonetheless, these products are accumulated using high-resolution (typically 32-bit) additions, an operation that dominates the arithmetic complexity of inference when using extreme quantization (e.g., binary weights). To further optimize inference, we propose a method that adapts neural networks to use low-resolution (8-bit) additions in the accumulators, achieving classification accuracy comparable to their 32-bit counterparts. We achieve resilience to low-resolution accumulation by inserting a cyclic activation layer, as well as an overflow penalty regularizer. We demonstrate the efficacy of our approach on both software and hardware platforms.
Transfer learning facilitates the training of task-specific classifiers using pre-trained models as feature extractors. We present a family of transferable adversarial attacks against such classifiers, generated without access to the classification head; we call these \emph{headless attacks}. We first demonstrate successful transfer attacks against a victim network using \textit{only} its feature extractor. This motivates the introduction of a label-blind adversarial attack. This transfer attack method does not require any information about the class-label space of the victim. Our attack lowers the accuracy of a ResNet18 trained on CIFAR10 by over 40\%.
Deep neural networks achieve state-of-the-art performance for a range of classification and inference tasks. However, the use of stochastic gradient descent combined with the nonconvexity of the underlying optimization problems renders parameter learning susceptible to initialization. To address this issue, a variety of methods that rely on random parameter initialization or knowledge distillation have been proposed in the past. In this paper, we propose FuseInit, a novel method to initialize shallower networks by fusing neighboring layers of deeper networks that are trained with random initialization. We develop theoretical results and efficient algorithms for mean-square error (MSE)-optimal fusion of neighboring dense-dense, convolutional-dense, and convolutional-convolutional layers. We show experiments for a range of classification and regression datasets, which suggest that deeper neural networks are less sensitive to initialization and shallower networks can perform better (sometimes as well as their deeper counterparts) if initialized with FuseInit.