Dependent Quantization (DQ) is a new quantization tool introduced in the Versatile Video Coding (VVC) standard. While it provides better rate-distortion calculation accuracy, it also increases the computational complexity and hardware cost compared to the widely used scalar quantization. To address this issue, this paper proposes a parallel-dependent quantization hardware architecture using Verilog HDL language. The architecture preprocesses the coefficients with a scalar quantizer and a high-frequency filter, and then further segments and processes the coefficients in parallel using the Viterbi algorithm. Additionally, the weight bit width of the rate-distortion calculation is reduced to decrease the quantization cycle and computational complexity. Finally, the final quantization of the TU is determined through sequential scanning and judging of the rate-distortion cost. Experimental results show that the proposed algorithm reduces the quantization cycle by an average of 56.96% compared to VVC’s reference platform VTM, with a Bjøntegaard delta bit rate (BDBR) loss of 1.03% and 1.05% under the Low-delay P and Random Access configurations, respectively. Verification on the AMD FPGA development platform demonstrates that the hardware implementation meets the quantization requirements for 1080P@60Hz video hardware encoding.

This paper proposes and evaluates machine learning (ML)-based compensation methods for the transmit (Tx) weight matrices of actual singular value decomposition (SVD)-multiple-input and multiple-output (MIMO) transmissions. These methods train ML models and compensate the Tx weight matrices by using a large amount of training data created from statistical distributions. Moreover, this paper proposes simplified channel metrics based on the channel quality of actual SVD-MIMO transmissions to evaluate compensation performance. The optimal parameters are determined from many ML parameters by using the metrics, and the metrics for this determination are evaluated. Finally, a comprehensive computer simulation shows that the optimal parameters improve performance by up to 7.0dB compared with the conventional method.

Edge computing, which has been gaining attention in recent years, has many advantages, such as reducing the load on the cloud, not being affected by the communication environment, and providing excellent security. Therefore, many researchers have attempted to implement neural networks, which are representative of machine learning in edge computing. Neural networks can be divided into inference and learning parts; however, there has been little research on implementing the learning component in edge computing in contrast to the inference part. This is because learning requires more memory and computation than inference, easily exceeding the limit of resources available for edge computing. To overcome this problem, this research focuses on the optimizer, which is the heart of learning. In this paper, we introduce our new optimizer, hardware-oriented logarithmic momentum estimation (Holmes), which incorporates new perspectives not found in existing optimizers in terms of characteristics and strengths of hardware. The performance of Holmes was evaluated by comparing it with other optimizers with respect to learning progress and convergence speed. Important aspects of hardware implementation, such as memory and operation requirements are also discussed. The results show that Holmes is a good match for edge computing with relatively low resource requirements and fast learning convergence. Holmes will help create an era in which advanced machine learning can be realized on edge computing.

The spectral envelope parameter is a significant speech parameter in the vocoder's quality. Recently, the Vector Quantized Variational AutoEncoder (VQ-VAE) is a state-of-the-art end-to-end quantization method based on the deep learning model. This paper proposed a new technique for improving the embedding space learning of VQ-VAE with the Generative Adversarial Network for quantizing the spectral envelope parameter, called VQ-VAE-EMGAN. In experiments, we designed the quantizer for the spectral envelope parameters of the WORLD vocoder extracted from the 16kHz speech waveform. As the results shown, the proposed technique reduced the Log Spectral Distortion (LSD) around 0.5dB and increased the PESQ by around 0.17 on average for four target bit operations compared to the conventional VQ-VAE.

In this letter, we propose a new color quantization method suppressing saturation decrease. In the proposed method, saturation-based weight and intensity-based weight are used so that vivid colors are selected as the representative colors preferentially. Experiments show that the proposed method tends to select vivid colors even if they occupy only a small area in the image.

This letter presents an efficient technique to reduce the computational complexity involved in training binary convolutional neural networks (BCNN). The BCNN training shall be conducted focusing on the optimization of the sign of each weight element rather than the exact value itself in convention; in which, the sign of an element is not likely to be flipped anymore after it has been updated to have such a large magnitude to be clipped out. The proposed technique does not update such elements that have been clipped out and eliminates the computations involved in their optimization accordingly. The complexity reduction by the proposed technique is as high as 25.52% in training the BCNN model for the CIFAR-10 classification task, while the accuracy is maintained without severe degradation.

In this letter, we propose a color quantization method based on saliency. In the proposed method, the salient colors are selected as representative colors preferentially by using saliency as weights. Through experiments, we verify the effectiveness of the proposed method.

Designing an optimum quantizer can be treated as the optimization problem of finding the quantization indices that minimize the quantization error. One solution to the optimization problem, DP quantization, is based on dynamic programming. Some applications, such as bit-depth scalable codec and tone mapping, require the construction of multiple quantizers with different quantization levels, for example, from 12bit/channel to 10bit/channel and 8bit/channel. Unfortunately, the above mentioned DP quantization optimizes the quantizer for just one quantization level. That is, it is unable to simultaneously optimize multiple quantizers. Therefore, when DP quantization is used to design multiple quantizers, there are many redundant computations in the optimization process. This paper proposes an extended DP quantization with a complexity reduction algorithm for the optimal design of multiple quantizers. Experiments show that the proposed algorithm reduces complexity by 20.8%, on average, compared to conventional DP quantization.

Quantization is an important technique for implementing convolutional neural networks on edge devices. Quantization often requires relearning, but relearning sometimes cannot be always be applied because of issues such as cost or privacy. In such cases, it is important to know the numerical precision required to maintain accuracy. We accurately simulate calculations on hardware and accurately measure the relationship between accuracy and numerical precision.

A high degree of freedom in spectral domain allows us to accommodate additional optical signal processing for wavelength division multiplexing in photonic analog-to-digital conversion. We experimentally verified a spectral compression to save a necessary bandwidth for soliton self-frequency shift based optical quantization through the cascade of the four-wave mixing based and the sum-frequency generation based spectral compression. This approach can realize 0.03 nm individual bandwidth correspond to save up to more than 85 percent of bandwidth for 7-bit optical quantization in C-band.

Many studies of deep neural networks have reported inference accelerators for improved energy efficiency. We propose methods for further improving energy efficiency while maintaining recognition accuracy, which were developed by the co-design of a filter-by-filter quantization scheme with variable bit precision and a hardware architecture that fully supports it. Filter-wise quantization reduces the average bit precision of weights, so execution times and energy consumption for inference are reduced in proportion to the total number of computations multiplied by the average bit precision of weights. The hardware utilization is also improved by a bit-parallel architecture suitable for granularly quantized bit precision of weights. We implement the proposed architecture on an FPGA and demonstrate that the execution cycles are reduced to 1/5.3 for ResNet-50 on ImageNet in comparison with a conventional method, while maintaining recognition accuracy.

This paper proposes a deep neural network (DNN) based framework to address the problem of vector quantization (VQ) for high-dimensional data. The main challenge of applying DNN to VQ is how to reduce the binary coding error of the auto-encoder when the distribution of the coding units is far from binary. To address this problem, three fine-tuning methods have been adopted: 1) adding Gaussian noise to the input of the coding layer, 2) forcing the output of the coding layer to be binary, 3) adding a non-binary penalty term to the loss function. These fine-tuning methods have been extensively evaluated on quantizing speech magnitude spectra. The results demonstrated that each of the methods is useful for improving the coding performance. When implemented for quantizing 968-dimensional speech spectra using only 18-bit, the DNN-based VQ framework achieved an averaged PESQ of about 2.09, which is far beyond the capability of conventional VQ methods.

Receiver implementations with very low quantization resolution will play an important role in 5G, as high precision quantization and signal processing are costly in terms of computational resources and chip area. Therefore, low resolution receivers with quasi optimum performance will be required to meet complexity and latency constraints. The Information Bottleneck method allows for a novel, information centric approach to design such receivers. The method was originally introduced by Naftali Tishby et al. and mostly used in the machine learning field so far. Interestingly, it can also be applied to build surprisingly good digital communication receivers which work fundamentally different than state-of-the-art receivers. Instead of minimizing the quantization error, receiver components with maximum preservation of relevant information for a given bit width can be designed. All signal processing in the resulting receivers is performed using only simple lookup operations. In this paper, we first provide a brief introduction to the design of receiver components with the Information Bottleneck method. We keep referring to decoding of low-density parity-check codes as a practical example. The focus of the paper lies on practical decoder implementations on a digital signal processor which illustrate the potential of the proposed technique. An Information Bottleneck decoder with 4bit message passing decoding is found to outperform 8bit implementations of the well-known min-sum decoder in terms of bit error rate and to perform extremely close to an 8bit belief propagation decoder, while offering considerably higher net decoding throughput than both conventional decoders.

We formulate the design of an optimal quantizer as an optimization problem that finds the quantization indices that minimize quantization error. As a solution of the optimization problem, an approach based on dynamic programming, which is called DP quantization, is proposed. It is observed that quantized signals do not always contain all kinds of signal values which can be represented with given bit-depth. This property is called amplitude sparseness. Because quantization is the amplitude discretization of signal value, amplitude sparseness is closely related to quantizer design. Signal values with zero frequency do not impact quantization error, so there is the potential to reduce the complexity of the optimal quantizer by not computing signal values that have zero frequency. However, conventional methods for DP quantization were not designed to consider amplitude sparseness, and so fail to reduce complexity. The proposed algorithm offers a reduced complexity optimal quantizer that minimizes quantization error while addressing amplitude sparseness. Experimental results show that the proposed algorithm can achieve complexity reduction over conventional DP quantization by 82.9 to 84.2% on average.

The spaces between words and paragraphs are popular places for embedding data in data hiding techniques for text documents. Due to the low redundancy in text documents, the payload is limited to be small. As each bit of data is independently inserted into specific spaces in conventional methods, a malicious party may be able to modify the data without causing serious visible distortions. In this paper, we regard a collection of space lengths as a one-dimensional feature vector and embed watermark into its frequency components. To keep the secrecy of the embedded information, a random permutation and dither modulation are introduced in the operation. Furthermore, robustness against additive noise is enhanced by controlling the payload. In the proposed method, through experiments, we evaluated the trade-off among payload, distortion, and robustness.

Generalized Minimum Distance (GMD) decoding is a well-known soft-decision decoding for linear codes. Previous research on GMD decoding focused mainly on unquantized AWGN channels with BPSK signaling for binary linear codes. In this paper, a study on the design of a 4-level uniform quantizer for GMD decoding is given. In addition, an extended version of a GMD decoding algorithm for a 4-level quantizer is proposed, and the effectiveness of the proposed decoding is shown by simulation.

In this paper, a study of a sufficient condition on the optimality of a decoded codeword of soft-decision decodings for binary linear codes is shown for a quantized case. A typical uniform 4-level quantizer for soft-decision decodings is employed for the analysis. Simulation results on the (64,42,8) Reed-Muller code indicates that the condition is effective for SN ratios at 3[dB] or higher for any iterative style optimum decodings.

Event-triggered control is a method that the control input is updated only when a certain triggering condition is satisfied. In networked control systems, quantization errors via A/D conversion should be considered. In this paper, a new method for quantized event-triggered control with switching triggering conditions is proposed. For a discrete-time linear system, we consider the problem of finding a state-feedback controller such that the closed-loop system is uniformly ultimately bounded in a certain ellipsoid. This problem is reduced to an LMI (Linear Matrix Inequality) optimization problem. The volume of the ellipsoid may be adjusted. The effectiveness of the proposed method is presented by a numerical example.

In this paper, a study on the design and implementation of uniform 4-level quantizers for soft-decision decodings for binary linear codes is shown. Simulation results on quantized Viterbi decoding with a 4-level quantizer for the (64,42,8) Reed-Muller code show that the optimum stepsize, which is derived from the cutoff rate, gives an almost optimum error performance. In addition, the simulation results show that the case where the number of optimum codewords is larger than the one for a received sequence causes non-negligible degradation on error performance at high SN ratios of Eb/N0.

Side match vector quantization (SMVQ) has been originally developed for image compression and is also useful for steganography. SMVQ requires to create its own state codebook for each block in both encoding and decoding phases. Since the conventional method for the state codebook generation is extremely time-consuming, this letter proposes a fast generation method. The proposed method is tens times faster than the conventional one without loss of perceptual visual quality.