Artificial intelligence (AI) has emerged as a promising technology to design efficient channel state information (CSI) feedback methods. The effectiveness of learned CSI compression has been proven experimentally in many previous works, but less attention was paid to the importance of scalability in practice to support variable sizes of feedback payloads. In this paper, we propose employing multi-stage vector quantization (VQ) to provide progressive variable-rate CSI compression. In particular, VQ is cascaded to find a series of low-dimensional CSI representations, which provides progressively refined approximations associated with variable feedback rates. In practice, deploying a dedicated model with a separate codebook optimized for each rate requirement is prominently challenging in the aspects of storage, on-device compiling, if needed, and model life-cycle management. To tackle this challenge, we develop a single network model with multi-stage codebooks, which is learned for various compression rates of interest jointly. Moreover, we propose employing sequential training to aid joint training by providing good initialization of network parameters with transfer learning. The simulation results shows the effectiveness of the proposed model and training method, which establish a promising design methodology of AI-enabled CSI compression.

Stopping at a red light at intersections is one of the main contributors to travel time delays in urban traffic. Drivers often stop as close as possible to their lead vehicle, which may cause additional time delay because they should wait before accelerating to secure a safe distance. In this paper, we present an infrastructure-based optimal spacing system that guides vehicles to maintain the optimal spacing when coming to a stop. To achieve this, we formulate an optimization problem to compute the optimal spacing between vehicles such that all vehicles can enter the intersection as soon as possible when the light turns green. To solve the problem efficiently, we decompose it into sub-problems involving only two vehicles and sequentially solve them. We validate through simulations that our approach effectively reduces travel time delay.

In multiple-input multiple-output (MIMO) communication systems, precoding is an essential component to increase the data rate by projecting the transmit signal to the channel space of good condition. Singular value decomposition (SVD) is often used to produce the precoding matrix consisting of the right-singular vectors of the channel matrix, which is known to maximize channel capacity for Gaussian inputs. However, for non-Gaussian inputs, it is desired for the precoder to maximize mutual information (MI) between the channel input and output directly, since the theoretical channel capacity is only achievable under Gaussian input assumption. In this paper, we investigate a data-driven approach to learn the MI-maximizing precoder for finite-alphabet inputs. In particular, we focus on training a single model capable of producing the optimal precoder for various MIMO system configurations regarding the number of antennas, the number of data streams, and the modulation order, so it can be deployed easily in practice without the overhead of storing and managing multiple models. In our simulation results, it is shown that the proposed learned precoder results in higher MI values and lower block error rates in various scenarios, compared to the conventional capacity-maximizing SVD-based precoder.

Large-scale matrix computations have become indispensable in artificial intelligence and scientific applications. It is of paramount importance to efficiently perform out-of-core computations that often entail an excessive amount of disk I/O. Unfortunately, however, most existing systems do not focus on disk I/O aspects and are vulnerable to performance degradation when the scale of input matrices and intermediate data grows large. To address this problem, we present a new out-of-core matrix computation system called PreVision. The PreVision system can achieve optimal buffer replacement by leveraging the deterministic characteristics of data access patterns, and it can also avoid redundant I/O operations by proactively evicting the pages that are no longer referenced. Through extensive evaluations, we demonstrate that PreVision outperforms the existing out-of-core matrix computation systems and significantly reduces disk I/O operations.

We propose a novel method for training a conditional generative adversarial network (CGAN) without the use of training data, called zero-shot learning of a CGAN (ZS-CGAN). Zero-shot learning of a conditional generator only needs a pre-trained discriminative (classification) model and does not need any training data. In particular, the conditional generator is trained to produce labeled synthetic samples whose characteristics mimic the original training data by using the statistics stored in the batch normalization layers of the pretrained model. We show the usefulness of ZS-CGAN in data-free quantization of deep neural networks. We achieved the state-of-the-art data-free network quantization of the ResNet and MobileNet classification models trained on the ImageNet dataset. Data-free quantization using ZS-CGAN showed a minimal loss in accuracy compared to that obtained by conventional data-dependent quantization.

We consider the optimization of deep convolutional neural networks (CNNs) such that they provide good performance while having reduced complexity if deployed on either conventional systems with spatial-domain convolution or lower-complexity systems designed for Winograd convolution. The proposed framework produces one compressed model whose convolutional filters can be made sparse either in the spatial domain or in the Winograd domain. Hence, the compressed model can be deployed universally on any platform, without need for re-training on the deployed platform. To get a better compression ratio, the sparse model is compressed in the spatial domain that has a fewer number of parameters. From our experiments, we obtain 24.2× and 47.7× compressed models for ResNet-18 and AlexNet trained on the ImageNet dataset, while their computational cost is also reduced by 4.5× and 5.1×, respectively.

In this paper, we propose a novel variable-rate learned image compression framework with a conditional autoencoder. Previous learning-based image compression methods mostly require training separate networks for different compression rates so they can yield compressed images of varying quality. In contrast, we train and deploy only one variable-rate image compression network implemented with a conditional autoencoder. We provide two rate control parameters, i.e., the Lagrange multiplier and the quantization bin size, which are given as conditioning variables to the network. Coarse rate adaptation to a target is performed by changing the Lagrange multiplier, while the rate can be further fine-tuned by adjusting the bin size used in quantizing the encoded representation. Our experimental results show that the proposed scheme provides a better rate-distortion trade-off than the traditional variable-rate image compression codecs such as JPEG2000 and BPG. Our model also shows comparable and sometimes better performance than the state-of-the-art learned image compression models that deploy multiple networks trained for varying rates.

Pilot-aided channel estimation has been popular for 4G wideband communication systems. Even though it is convenient, its limitation comes from the fact that the performance is bounded by the density of pilot symbols. Increasing pilot symbols improves channel estimation quality, but it also hurts bandwidth efficiency. In this paper, we advocate data-aided channel estimation for 5G. It is potentially the only solution to improve channel estimation quality without increasing pilot density. In particular, we aim to answer two fundamental questions for data-aided channel estimation: (1) What is the optimal scheme for data-aided channel estimation? (2) How can we make the optimal scheme practical by reducing its complexity? We present how we tackle these problems of data-aided channel estimation and show that the proposed iterative scheme yields significant gain in long term evolution (LTE) signal demodulation.

In this paper, we propose a novel method of reducing the complexity of a pilot-based two-dimensional linear minimum mean square error (2D LMMSE) channel estimation scheme for orthogonal frequency division multiplexing (OFDM) systems. We identify that the 2D LMMSE channel estimation method aided by pilots embedded in a two-dimensional OFDM resource grid can be decoupled into three steps: (a) pilot denoising, (b) interpolation in each of OFDM symbols including pilots, and (c) interpolation across OFDM symbols. In the denoising step, we process pilots only for noise reduction. In the following interpolation steps, we utilize two-dimensional channel correlation across subcarriers (frequency) and across OFDM symbols (time) in order to get channel estimates of interest. Under the assumption that the frequency and time correlations are disjoint and the channel correlation can be represented by the product of them, we show that the channel interpolation can be performed in two steps along the frequency domain first and then along the time domain. By taking advantage of this separation, we can reduce the complexity of 2D LMMSE considerably while still achieving the optimal performance of it.

In the presence of co-channel interference in cellular networks, interference mitigation by detecting the desired signal jointly with the interference promises considerable gain over the conventional way of handling the interference as colored Gaussian. Even though such interference-aware detection can improve the performance, it requires some information on the interference. In particular, the modulation format of the interference has to be classified to this end, when it is not signaled by the network explicitly. This paper investigates interference modulation classification methods for interference-aware joint detection. We propose an iterative interference modulation classification algorithm that utilizes the decoded information of the desired signal in order to cancel the desired signal from the received signal. After the cancellation, the remaining signal can be treated as interference plus noise so that we can classify the modulation format of the interference at reduced complexity with small performance loss due to decoding errors.

This paper considers the problem of mismatched hypothesis testing, where approximate likelihood functions are used instead of true likelihood functions. Given a hypothesis testing problem, the maximum likelihood (ML) solution is known to be optimal when true likelihood functions are used, but the optimality does not hold anymore if mismatched approximate likelihood functions are employed instead, in order to reduce computational complexity, for instance. In this paper, we investigate the mismatched ML framework using approximate likelihood functions, while the mismatches between the true and the approximate likelihood functions are corrected by additive compensating constants. The probability of error of this mismatched hypothesis testing is analyzed asymptotically, assuming a large number of samples, and the compensating constants that maximize the error exponent are established. The general results on the mismatched hypothesis testing are then utilized in designing and optimizing a digital modulation classifier with low complexity.

The effectiveness of application-layer coding in a system with a large number of users is considered. The end users encode data packets before transmitting them. The effect of additional packets on the system performance is twofold: (i) additional packets increase offered load, which results in higher drop probability, and (ii) some of dropped packets can be recovered at the receivers after decoding. It is argued that the space of all systems can be partitioned into two regions where coding is beneficial and detrimental, respectively. In particular, the paper establishes an asymptotic regime that contains the boundary between these two regions. On the boundary, systems with and without coding have the same performance. Informally, our results indicate that application-layer coding improves the performance only in systems with low loss probabilities (without coding), and employing such coding in systems with high loss probabilities only degrades the performance.

The effectiveness of application-layer coding in a system with a large number of users is considered. The end users encode data packets before transmitting them. The effect of additional packets on the system performance is twofold: (i) additional packets increase the offered load, which results in higher drop probability, and (ii) some of dropped packets can be recovered at the receivers after decoding. The paper establishes an asymptotic regime in which systems with and without coding have the same performance. The space of all systems is partitioned into two regions where coding is beneficial and detrimental, respectively. Informally, it is argued that application-layer coding improves the performance only in systems with low loss probabilities (without coding), and employing such coding in systems with high loss probabilities only degrades the performance.

In this paper, we propose a fast method for separating reflection components using a single color image. We first propose a specular-free two-band image that is a specularity-invariant color image representation. Reflection components separation is achieved by comparing local ratios at each pixel and making those ratios equal in an iterative framework. The proposed method is very fast and shows reasonable results for textured indoor/outdoor images.