Search

This work presents the first silicon-validated dedicated EGM-to-ECG (G2C) processor, dubbed e-G2C, featuring continuous lightweight anomaly detection, event-driven coarse/precise conversion, and on-chip adaptation. e-G2C utilizes neural network (NN) based G2C conversion and integrates 1) an architecture supporting anomaly detection and coarse/precise conversion via time multiplexing to balance the effectiveness and power, 2) an algorithm-hardware co-designed vector-wise sparsity resulting in a 1.6-1.7$\times$ speedup, 3) hybrid dataflows for enhancing near 100% utilization for normal/depth-wise(DW)/point-wise(PW) convolutions (Convs), and 4) an on-chip detection threshold adaptation engine for continuous effectiveness. The achieved 0.14-8.31 $\mu$J/inference energy efficiency outperforms prior arts under similar complexity, promising real-time detection/conversion and possibly life-critical interventions, Comment: Accepted by 2022 IEEE Symposium on VLSI Technology and Circuits (VLSI Technology and Circuits)

K-means defines one of the most employed centroid-based clustering algorithms with performances tied to the data's embedding. Intricate data embeddings have been designed to push $K$-means performances at the cost of reduced theoretical guarantees and interpretability of the results. Instead, we propose preserving the intrinsic data space and augment K-means with a similarity measure invariant to non-rigid transformations. This enables (i) the reduction of intrinsic nuisances associated with the data, reducing the complexity of the clustering task and increasing performances and producing state-of-the-art results, (ii) clustering in the input space of the data, leading to a fully interpretable clustering algorithm, and (iii) the benefit of convergence guarantees., Comment: arXiv admin note: substantial text overlap with arXiv:2012.09743

There exists a gap in terms of the signals provided by pacemakers (i.e., intracardiac electrogram (EGM)) and the signals doctors use (i.e., 12-lead electrocardiogram (ECG)) to diagnose abnormal rhythms. Therefore, the former, even if remotely transmitted, are not sufficient for doctors to provide a precise diagnosis, let alone make a timely intervention. To close this gap and make a heuristic step towards real-time critical intervention in instant response to irregular and infrequent ventricular rhythms, we propose a new framework dubbed RT-RCG to automatically search for (1) efficient Deep Neural Network (DNN) structures and then (2)corresponding accelerators, to enable Real-Time and high-quality Reconstruction of ECG signals from EGM signals. Specifically, RT-RCG proposes a new DNN search space tailored for ECG reconstruction from EGM signals, and incorporates a differentiable acceleration search (DAS) engine to efficiently navigate over the large and discrete accelerator design space to generate optimized accelerators. Extensive experiments and ablation studies under various settings consistently validate the effectiveness of our RT-RCG. To the best of our knowledge, RT-RCG is the first to leverage neural architecture search (NAS) to simultaneously tackle both reconstruction efficacy and efficiency., Comment: JETC Special issue on Hardware-Aware Learning for Medical Applications (Waiting assignment to batch)

We design an interpretable clustering algorithm aware of the nonlinear structure of image manifolds. Our approach leverages the interpretability of $K$-means applied in the image space while addressing its clustering performance issues. Specifically, we develop a measure of similarity between images and centroids that encompasses a general class of deformations: diffeomorphisms, rendering the clustering invariant to them. Our work leverages the Thin-Plate Spline interpolation technique to efficiently learn diffeomorphisms best characterizing the image manifolds. Extensive numerical simulations show that our approach competes with state-of-the-art methods on various datasets.

The gold standard of a wireless network is that the throughput increases linearly with the density of access points (APs). However, such a linear throughput gain is suspended in the 5G mmWave network mainly due to the short communication distances in mmWave bands and the dense deployments of mmWave APs. As being operated in the interference-limited regime, the aggregate interference resulted from the increasing mmWave APs will gradually become the network performance bottleneck, which leads to the saturation of the throughput. In this paper, we propose to overcome the densification plateau of a mmWave network by employing spatial multiplexing at APs. To study the effect of spatial multiplexing on mmWave networks, we first derive the coverage probability as a function of spatial multiplexing gain. The fixed-rate coding scheme is then used to provide the network throughput. We also introduce the concept of densification gain to capture the improvement in network throughput achieved through the AP densification. Our results indicate that without the spatial multiplexing at APs, the throughput of mmWave networks will reach the plateau when the density of APs becomes sufficiently large. By enabling the spatial multiplexing at APs, however, the mmWave network can continuously harvest the throughput gains as the number of APs grows. By deriving the upper bound for the throughput of mmWave networks, we quantify the potential throughput improvement as the spatial multiplexing gain increases. However, our numerical results show that such a potential throughput gain cannot be explored by the fixed-rate coding scheme. We then demonstrate the necessity for deploying the multi-rate coding scheme in mmWave networks, especially when the spatial multiplexing gain at mmWave APs is large., Comment: Submitted to IEEE Transactions on Wireless Communications

We propose the Nonlinear Regression Convolutional Encoder-Decoder (NRCED), a novel framework for mapping a multivariate input to a multivariate output. In particular, we implement our algorithm within the scope of 12-lead surface electrocardiogram (ECG) reconstruction from intracardiac electrograms (EGM) and vice versa. The goal of performing this task is to allow for improved point-of-care monitoring of patients with an implanted device to treat cardiac pathologies. We will achieve this goal with 12-lead ECG reconstruction and by providing a new diagnostic tool for classifying atypical heartbeats. The algorithm is evaluated on a dataset retroactively collected from 14 patients. Correlation coefficients calculated between the reconstructed and the actual ECG show that the proposed NRCED method represents an efficient, accurate, and superior way to synthesize a 12-lead ECG. We can also achieve the same reconstruction accuracy with only one EGM lead as input. We also tested the model in a non-patient specific way and saw a reasonable correlation coefficient. The model was also executed in the reverse direction to produce EGM signals from a 12-lead ECG and found that the correlation was comparable to the forward direction. Lastly, we analyzed the features learned in the model and determined that the model learns an overcomplete basis of our 12-lead ECG space. We then use this basis of features to create a new diagnostic tool for identifying atypical and diseased heartbeats. This resulted in a ROC curve with an associated area under the curve value of 0.98, demonstrating excellent discrimination between the two classes.

A big mystery in deep learning continues to be the ability of methods to generalize when the number of model parameters is larger than the number of training examples. In this work, we take a step towards a better understanding of the underlying phenomena of Deep Autoencoders (AEs), a mainstream deep learning solution for learning compressed, interpretable, and structured data representations. In particular, we interpret how AEs approximate the data manifold by exploiting their continuous piecewise affine structure. Our reformulation of AEs provides new insights into their mapping, reconstruction guarantees, as well as an interpretation of commonly used regularization techniques. We leverage these findings to derive two new regularizations that enable AEs to capture the inherent symmetry in the data. Our regularizations leverage recent advances in the group of transformation learning to enable AEs to better approximate the data manifold without explicitly defining the group underlying the manifold. Under the assumption that the symmetry of the data can be explained by a Lie group, we prove that the regularizations ensure the generalization of the corresponding AEs. A range of experimental evaluations demonstrate that our methods outperform other state-of-the-art regularization techniques.

The operating range of a single millimeter wave (mmWave) access point (AP) is small due to the high path loss and blockage issues of the frequency band. To achieve the coverage similar to conventional sub-6GHz networks, the ultra-dense deployments of APs are required by the mmWave network. In general, the mmWave APs can be categorized into backhaul-connected APs and relay APs. Though the spatial distribution of backhaul-connected APs can be captured by the Poison point process (PPP), the desired locations of relay APs depend on the transmission protocol operated in the mmWave network. In this paper, we consider modeling the topology of mmWave AP network by incorporating the multihop protocol. We first derive the topology of AP network with the spatial multiplexing disabled for each transmission hop. Then we analyze the topology when the spatial multiplexing is enabled at the mmWave APs. To derive the network throughput, we first quantify the improvement in latency and the degradation of coverage probability with the increase of spatial multiplexing gain at mmWave APs. Then we show the impact of spatial multiplexing on the throughput for the ultra-dense mmWave AP network., Comment: submitted to ISIT 2020

We study the geometry of deep (neural) networks (DNs) with piecewise affine and convex nonlinearities. The layers of such DNs have been shown to be {\em max-affine spline operators} (MASOs) that partition their input space and apply a region-dependent affine mapping to their input to produce their output. We demonstrate that each MASO layer's input space partitioning corresponds to a {\em power diagram} (an extension of the classical Voronoi tiling) with a number of regions that grows exponentially with respect to the number of units (neurons). We further show that a composition of MASO layers (e.g., the entire DN) produces a progressively subdivided power diagram and provide its analytical form. The subdivision process constrains the affine maps on the (exponentially many) power diagram regions to greatly reduce their complexity. For classification problems, we obtain a formula for a MASO DN's decision boundary in the input space plus a measure of its curvature that depends on the DN's nonlinearities, weights, and architecture. Numerous numerical experiments support and extend our theoretical results.

We define a metric, mutual information in frequency (MI-in-frequency), to detect and quantify the statistical dependence between different frequency components in the data, referred to as cross-frequency coupling and apply it to electrophysiological recordings from the brain to infer cross-frequency coupling. The current metrics used to quantify the cross-frequency coupling in neuroscience cannot detect if two frequency components in non-Gaussian brain recordings are statistically independent or not. Our MI-in-frequency metric, based on Shannon's mutual information between the Cramer's representation of stochastic processes, overcomes this shortcoming and can detect statistical dependence in frequency between non-Gaussian signals. We then describe two data-driven estimators of MI-in-frequency: one based on kernel density estimation and the other based on the nearest neighbor algorithm and validate their performance on simulated data. We then use MI-in-frequency to estimate mutual information between two data streams that are dependent across time, without making any parametric model assumptions. Finally, we use the MI-in- frequency metric to investigate the cross-frequency coupling in seizure onset zone from electrocorticographic recordings during seizures. The inferred cross-frequency coupling characteristics are essential to optimize the spatial and spectral parameters of electrical stimulation based treatments of epilepsy., Comment: This paper is accepted for publication in IEEE Transactions on Signal Processing and contains 15 pages, 9 figures and 1 table

We consider the problem of estimating mutual information between dependent data, an important problem in many science and engineering applications. We propose a data-driven, non-parametric estimator of mutual information in this paper. The main novelty of our solution lies in transforming the data to frequency domain to make the problem tractable. We define a novel metric--mutual information in frequency--to detect and quantify the dependence between two random processes across frequency using Cram\'{e}r's spectral representation. Our solution calculates mutual information as a function of frequency to estimate the mutual information between the dependent data over time. We validate its performance on linear and nonlinear models. In addition, mutual information in frequency estimated as a part of our solution can also be used to infer cross-frequency coupling in the data., Comment: Submitted to International Symposium on Information Theory (ISIT) 2017. 5 pages, 6 figures

We present a sparse and invariant representation with low asymptotic complexity for robust unsupervised transient and onset zone detection in noisy environments. This unsupervised approach is based on wavelet transforms and leverages the scattering network from Mallat et al. by deriving frequency invariance. This frequency invariance is a key concept to enforce robust representations of transients in presence of possible frequency shifts and perturbations occurring in the original signal. Implementation details as well as complexity analysis are provided in addition of the theoretical framework and the invariance properties. In this work, our primary application consists of predicting the onset of seizure in epileptic patients from subdural recordings as well as detecting inter-ictal spikes., Comment: 10 pages + 1 reference page

Integer-forcing (IF) precoding, also known as downlink IF, is a promising new approach for communication over multiple-input multiple-output (MIMO) broadcast channels. Inspired by the integer-forcing linear receiver for multiple-access channels, it generalizes linear precoding by inducing an effective channel matrix that is approximately integer, rather than approximately identity. Combined with lattice encoding and a pre-inversion of the channel matrix at the transmitter, the scheme has the potential to outperform any linear precoding scheme, despite enjoying similar low complexity. In this paper, a specific IF precoding scheme, called diagonally-scaled exact IF (DIF), is proposed and shown to achieve maximum spatial multiplexing gain. For the special case of two receivers, in the high SNR regime, an optimal choice of parameters is derived analytically, leading to an almost closed-form expression for the achievable sum rate. In particular, it is shown that the gap to the sum capacity is upper bounded by 0.27 bits for any channel realization. For general SNR, a regularized version of DIF (RDIF) is proposed. Numerical results for two receivers under Rayleigh fading show that RDIF can achieve performance superior to optimal linear precoding and very close to the sum capacity., Comment: 13 pages, 5 figures. To appear at the IEEE Transactions on Wireless Communications

We propose two low-complexity lattice code constructions that have competitive coding and shaping gains. The first construction, named systematic Voronoi shaping, maps short blocks of integers to the dithered Voronoi integers, which are dithered integers that are uniformly distributed over the Voronoi region of a low-dimensional shaping lattice. Then, these dithered Voronoi integers are encoded using a high-dimensional lattice retaining the same shaping and coding gains of low and high-dimensional lattices. A drawback to this construction is that there is no isomorphism between the underlying message and the lattice code, preventing its use in applications such as compute-and- forward. Therefore we propose a second construction, called mixed nested lattice codes, in which a high-dimensional coding lattice is nested inside a concatenation of low-dimensional shaping lattices. This construction not only retains the same shaping/coding gains as first construction but also provides the desired algebraic structure. We numerically study these methods, for point-to-point channels as well as compute-and-forward using low-density lattice codes (LDLCs) as coding lattices and E8 and Barnes-Wall as shaping lattices. Numerical results indicate a shaping gain of up to 0.86 dB, compared to the state-of-the-art of 0.4 dB; furthermore, the proposed method has lower complexity than state-of-the-art approaches., Comment: 13 pages

In this paper, we developed a model-based and a data-driven estimator for directed information (DI) to infer the causal connectivity graph between electrocorticographic (ECoG) signals recorded from brain and to identify the seizure onset zone (SOZ) in epileptic patients. Directed information, an information theoretic quantity, is a general metric to infer causal connectivity between time-series and is not restricted to a particular class of models unlike the popular metrics based on Granger causality or transfer entropy. The proposed estimators are shown to be almost surely convergent. Causal connectivity between ECoG electrodes in five epileptic patients is inferred using the proposed DI estimators, after validating their performance on simulated data. We then proposed a model-based and a data-driven SOZ identification algorithm to identify SOZ from the causal connectivity inferred using model-based and data-driven DI estimators respectively. The data-driven SOZ identification outperforms the model-based SOZ identification algorithm when benchmarked against visual analysis by neurologist, the current clinical gold standard. The causal connectivity analysis presented here is the first step towards developing novel non-surgical treatments for epilepsy., Comment: This paper is accepted for publication in IEEE Journal of Selected Topics in Signal Processing, special issue on Advanced Signal Processing in Brain Networks, October 2016. 16 pages, 11 figures and 2 tables

We derive the optimal signaling for a multiple input multiple output (MIMO) full-duplex two-way channel under the imperfect transmit front-end chain. We characterize the two-way rates of the channel by using a game-theoretical approach, where we focus on the Pareto boundary of the achievable rate region and Nash equilibia (NE). For a MISO full-duplex two-way channel, we prove that beamforming is an optimal transmission strategy which can achieve any point on the Pareto boundary. Furthermore, we present a closed-form expression for the optimal beamforming weights. In our numerical examples we quantify gains in the achievable rates of the proposed beamforming over the zero-forcing beamforming. For a general MIMO full-duplex channel, we establish the existence of NE and present a condition for the uniqueness of NE. We then propose an iterative water-filling algorithm which is capable of reaching NE. Through simulations the threshold of the self-interference level is found, below which the full-duplex NE outperforms the half-duplex TDMA., Comment: submitted to IEEE transactions on wireless communications

We model the self-interference in a multiple input single output (MISO) full-duplex two-way channel and evaluate the achievable rate region. We formulate the boundary of the achievable rate region termed as the Pareto boundary by a family of coupled, non-convex optimization problems. Our main contribution is decoupling and reformulating the original non-convex optimization problems to a family of convex semidefinite programming problems. For a MISO full-duplex two-way channel, we prove that beamforming is an optimal transmission strategy which can achieve any point on the Pareto boundary. Furthermore, we present a closed-form expression for the optimal beamforming weights. In our numerical examples we quantify gains in the achievable rates of the proposed beamforming over the zero-forcing beamforming., Comment: To appear in IEEE ICC 2015, London, UK

Recent wireless testbed implementations have proven that full-duplex communication is in fact possible and can outperform half-duplex systems. Many of these implementations modify existing half-duplex systems to operate in full-duplex. To realize the full potential of full-duplex, radios need to be designed with self-interference in mind. In our work, we use an experimental setup with a patch antenna prototype to characterize the self-interference channel between two radios. In doing so, we form an analytical model to design analog baseband cancellation techniques. We show that our cancellation scheme can provide up to 10 dB improved signal strength, 2.5 bps/Hz increase in rate, and a 10,000 improvement in BER as compared to the RF only cancellation provided by the patch antenna., Comment: 5 pages, 8 figures, to appear in 2013 Asilomar Conference proceedings

Recent wireless testbed implementations have proven that full-duplex communication is in fact possible and can outperform half-duplex systems. Many of these implementations modify existing half-duplex systems to operate in full-duplex. To realize the full potential of full-duplex, radios need to be designed with self-interference in mind. In our work, we use a novel patch antenna prototype in an experimental setup to characterize the self-interference channel between transmit and receive radios. We derive an equivalent analytical baseband model and propose analog baseband cancellation techniques to complement the RF cancellation provided by the patch antenna prototype. Our results show that a wide bandwidth, moderate isolation scheme achieves up to 2.4 bps/Hz higher achievable rate than a narrow bandwidth, high isolation scheme. Furthermore, the analog baseband cancellation yields a 10-10,000 improvement in BER over RF only cancellation., Comment: 22 pages, 12 figures, submitted to IEEE JSAC special issue on Full-Duplex

Femtocells are an emerging technology aimed at providing gains to both network operators and end-users. These gains come at a cost of increased interference, specifically the cross network interference between the macrocell and femtocell networks. This interference is one of the main performance limiting factors in allowing an underlaid femtocell network to share the spectrum with the cellular network. To manage this interference, we first propose a femtocell architecture that orthog- onally partitions the network bandwidth between the macrocell and femtocell networks. This scheme eliminates the cross network interference thus giving the femtocells more freedom over their use of the spectrum. Specifically, no interference constraint is imposed by the cellular network allowing femto users to transmit at a constant power on randomly selected channels. Although simple, this scheme is enough to give gains up to 200% in sum rate. We then propose a second architecture where both networks share the bandwidth simultaneously. A femtocell power control scheme that relies on minimal coordination with the macrocell base station is used in conjunction with an interference sensing channel assignment mechanism. These two schemes together yield sum rate gains up to 200%. We then develop a technique for macro users to join a nearby femtocell and share a common chan- nel with a femtocell user through the use of successive interfer- ence cancellation. By adding this mechanism to the power control and channel assignment schemes, we show sum rate gains over 300% and up to 90% power savings for macrocell users., Comment: 9 pages, 8 figures, Accepted to the International Journal of Advances in Engineering Sciences and Applied Mathematics special issue on Multi-Terminal Information Theory

In an attempt to utilize spectrum resources more efficiently, protocols sharing licensed spectrum with unlicensed users are receiving increased attention. From the perspective of cellular networks, spectrum underutilization makes spatial reuse a feasible complement to existing standards. Interference management is a major component in designing these schemes as it is critical that licensed users maintain their expected quality of service. We develop a distributed dynamic spectrum protocol in which ad-hoc device-to-device users opportunistically access the spectrum actively in use by cellular users. First, channel gain estimates are used to set feasible transmit powers for device-to-device users that keeps the interference they cause within the allowed interference temperature. Then network information is distributed by route discovery packets in a random access manner to help establish either a single-hop or multi-hop route between two device-to-device users. We show that network information in the discovery packet can decrease the failure rate of the route discovery and reduce the number of necessary transmissions to find a route. Using the found route, we show that two device-to-device users can communicate with a low probability of outage while only minimally affecting the cellular network, and can achieve significant power savings when communicating directly with each other instead of utilizing the cellular base station., Comment: 11 pages, 11 figues, Published in IEEE Trans. Wireless Communications

We carry out a comprehensive study of the resource cost of averaging consensus in wireless networks. Most previous approaches suppose a graphical network, which abstracts away crucial features of the wireless medium, and measure resource consumption only in terms of the total number of transmissions required to achieve consensus. Under a path-loss dominated model, we study the resource requirements of consensus with respect to three wireless-appropriate metrics: total transmit energy, elapsed time, and time-bandwidth product. First we characterize the performance of several popular gossip algorithms, showing that they may be order-optimal with respect to transmit energy but are strictly suboptimal with respect to elapsed time and time-bandwidth product. Further, we propose a new consensus scheme, termed hierarchical averaging, and show that it is nearly order-optimal with respect to all three metrics. Finally, we examine the effects of quantization, showing that hierarchical averaging provides a nearly order-optimal tradeoff between resource consumption and quantization error., Comment: 12 pages, 3 figures, to appear in IEEE Journal Selected Topics in Signal Processing, April 2013

We examine the benefits of user cooperation under compute-and-forward. Much like in network coding, receivers in a compute-and-forward network recover finite-field linear combinations of transmitters' messages. Recovery is enabled by linear codes: transmitters map messages to a linear codebook, and receivers attempt to decode the incoming superposition of signals to an integer combination of codewords. However, the achievable computation rates are low if channel gains do not correspond to a suitable linear combination. In response to this challenge, we propose a cooperative approach to compute-and-forward. We devise a lattice-coding approach to block Markov encoding with which we construct a decode-and-forward style computation strategy. Transmitters broadcast lattice codewords, decode each other's messages, and then cooperatively transmit resolution information to aid receivers in decoding the integer combinations. Using our strategy, we show that cooperation offers a significant improvement both in the achievable computation rate and in the diversity-multiplexing tradeoff., Comment: submitted to IEEE Transactions on Information Theory

In this paper, we present the results from over-the-air experiments of a complete implementation of an amplify and forward cooperative communications system. Our custom OFDM-based physical layer uses a distributed version of the Alamouti block code, where the relay sends one branch of Alamouti encoded symbols. First we show analytically and experimentally that amplify and forward protocols are unaffected by carrier frequency offsets at the relay. This result allows us to use a conventional Alamouti receiver without change for the distributed relay system. Our full system implementation shows gains up to 5.5dB in peak power constrained networks. Thus, we can conclusively state that even the simplest form of relaying can lead to significant gains in practical implementations.