Block Scaling Research Articles

Estimating Classification Accuracy for Unlabeled Datasets Based on Block Scaling

This paper proposes an approach called block scaling quality (BSQ) for estimating the prediction accuracy of a deep network model. The basic operation perturbs the input spectrogram by multiplying all values within a block by , where is equal to 0 in the experiments. The ratio of perturbed spectrograms that have different prediction labels than the original spectrogram to the total number of perturbed spectrograms indicates how much of the spectrogram is crucial for the prediction. Thus, this ratio is inversely correlated with the accuracy of the dataset. The BSQ approach demonstrates satisfactory estimation accuracy in experiments when compared with various other approaches. When using only the Jamendo and FMA datasets, the estimation accuracy experiences an average error of 4.9% and 1.8%, respectively. Moreover, the BSQ approach holds advantages over some of the comparison counterparts. Overall, it presents a promising approach for estimating the accuracy of a deep network model.

Boosting adversarial attacks with future momentum and future transformation

The transferability of adversarial examples under the black-box attack setting has attracted extensive attention from the community. Advanced optimization algorithms are one of the most successful ways to improve the transferability among all methods proposed recently. However, existing advanced optimization algorithms either can only slightly enhance the transferability of adversarial examples or take a large amount of computation time. We propose the future momentum and future transformation (FMFT) method to balance the transferability and computation overhead. The FMFT method incorporates two parts, future momentum (FM) and future transformation (FT). FM is inspired by the looking ahead property and updates adversarial examples with the future N-th step momentum during each iteration. FT, on the other hand, makes use of the input transformations during the progress of future momentum calculation to obtain a more robust gradient and reduce computation overhead. Additionally, we propose a new input transformation called random block scaling, which divides the image into various blocks and then scales the blocks differently. Empirical evaluations on the standard ImageNet dataset demonstrate the superiority of our FMFT.

NITI: Training Integer Neural Networks Using Integer-Only Arithmetic

Low bitwidth integer arithmetic has been widely adopted in hardware implementations of deep neural network inference applications. However, despite the promised energy-efficiency improvements demanding edge applications, the use of low bitwidth integer arithmetic for neural network training remains limited. Unlike inference, training demands high dynamic range and numerical accuracy for high quality results, making the use of low-bitwidth integer arithmetic particularly challenging. To address this challenge, we present a novel neural network training framework called NITI that exclusively utilizes low bitwidth integer arithmetic. NITI stores all parameters and accumulates intermediate values as 8-bit integers while using no more than 5 bits for gradients. To provide the necessary dynamic range during the training process, a per-layer block scaling exponentiation scheme is utilized. By deeply integrating with the rounding procedures and integer entropy loss calculation, the proposed scaling scheme incurs only minimal overhead in terms of storage and additional computation. Furthermore, a hardware-efficient pseudo-stochastic rounding scheme that eliminates the need for external random number generation is proposed to facilitate conversion from wider intermediate arithmetic results to lower precision for storage. Since NITI operates only with standard 8-bit integer arithmetic and storage, it is possible to accelerate it using existing low bitwidth operators originally developed for inference in commodity accelerators. To demonstrate this, an open-source software implementation of end-to-end training, using native 8-bit integer operations in modern GPUs is presented. In addition, experiments have been conducted on an FPGA-based training accelerator to evaluate the hardware advantage of NITI. When compared with an equivalent training setup implemented with floating point storage and arithmetic, NITI has no accuracy degradation on the MNIST and CIFAR10 datasets. On ImageNet, NITI achieves similar accuracy as state-of-the-art integer training frameworks without relying on full-precision floating-point first and last layers.

Pre-processing ensembles with response oriented sequential alternation calibration (PROSAC): A step towards ending the pre-processing search and optimization quest for near-infrared spectral modelling

Ensemble pre-processing is emerging as a potential tool to avoid the tiring pre-processing selection and optimization task in near-infrared (NIR) spectral modelling. Furthermore, differently pre-processed data may carry complementary information, hence, ensemble pre-processing may represent the best suited modelling option to extract all the useful information from differently pre-processed data. Recently, multi-block techniques such as sequential (SPORT) and parallel (PORTO) orthogonalized partial least squares regression were proposed to extract complementary information present in differently pre-processed data. Although such multi-block techniques allowed efficient modelling of differently pre-processed data blocks, depending on the approach, challenges related to choosing block order, parameter tuning, block scaling and optimization time requirements still must be dealt with. To cope with such issues, the present study proposes the use of a recently developed faster, block order independent and scale independent, multi-block data modelling technique called response-oriented sequential alternation (ROSA) to process the multi-block data generated by differently pre-processing the same NIR data. This new method is called PROSAC, i.e., pre-processing ensembles with ROSA calibration. The potential of the approach is demonstrated on five real NIR spectral datasets. Furthermore, as baselines for comparison, partial least squares regression was done on individually pre-processed data sets, and using two multi-block pre-processing fusion approaches, i.e., SPORT and PORTO. The ensemble pre-processing with ROSA achieved either better performance compared to the baseline methods or achieved comparable performance without the need to worry about the pre-processing order, the scaling of data after pre-processing and optimization time requirements. PROSAC can be considered as a general tool for the ensemble pre-processing for NIR data modelling. • The method was compared with recent novel multi-block pre-processing ensemble tools. • The method is fast, order and block scale independent thus complements pre-processing ensembles. • The method was evaluated on five NIR spectroscopy datasets. • The method eliminates the need for pre-processing selection.

Fast Power Series Solution of Large 3-D Electrodynamic Integral Equation for PEC Scatterers

This paper presents a new fast power series solution method to solve the Hierarchal Method of Moment (MoM) matrix for a large complex, perfectly electric conducting (PEC) 3D structures. The proposed power series solution converges in just two (2) iterations which is faster than the conventional fast solver–based iterative solution. The method is purely algebraic in nature and, as such applicable to existing conventional methods. The method uses regular fast solver Hierarchal Matrix (H-Matrix) and can also be applied to Multilevel Fast Multipole Method Algorithm (MLFMA). In the proposed method, we use the scaling of the symmetric near-field matrix to develop a diagonally dominant overall matrix to enable a power series solution. Left and right block scaling coefficients are required for scaling near-field blocks to diagonal blocks using Schur’s complement method. However, only the right-hand scaling coefficients are computed for symmetric near-field matrix leading to saving of computation time and memory. Due to symmetric property, the left side-block scaling coefficients are just the transpose of the right-scaling blocks. Next, the near-field blocks are replaced by scaled near-field diagonal blocks. Now the scaled near-field blocks in combination with far-field and scaling coefficients are subjected to power series solution terminating after only two terms. As all the operations are performed on the near-field blocks, the complexity of scaling coefficient computation is retained as O(N)O⁢(N). The power series solution only involves the matrix-vector product of the far-field, scaling coefficients blocks, and inverse of scaled near-field blocks. Hence, the solution cost remains O(NlogN)O⁢(N⁢l⁢o⁢g⁢N). Several numerical results are presented to validate the efficiency and robustness of the proposed numerical method.

High-Precision Binary Object Detector Based on a BSF-XNOR Convolutional Layer

Recently, building an efficient and robust model for object detection has attracted the attention of the vision community. Although binary networks have a fast inference speed, they cannot be used directly on mobile devices such as unmanned aerial vehicles (UAVs) because of their low detection accuracy. Different from improving the detection accuracy of a binary network by adjusting the network structure or adjusting the update gradient, we propose an improved binary neural network based on the block scaling factor XNOR (BSF-XNOR) convolutional layer. In addition, we propose a two-level densely connected network structure, which further enhances the network layer's feature representation capabilities. Experiments using the TensorFlow framework prove the effectiveness of our algorithm in improving accuracy. Compared with the original standard XNOR network, the mean average precision (mAP) detected by our algorithm on the PASCAL VOC dataset was improved. The experimental results on the VisDrone2019 UAV dataset confirm that our method achieves a better balance between speed and accuracy than previous methods. Our algorithm aims to guide and deploy high-precision binary networks on the embedded device and solves the problem of low-precision binary networks.

RoIP Compression Method in Burst Noise Upstream Environment for CATV Network

A hybrid fiber coaxial (HFC) network uses both optical fibers and coaxial cables for cable TV networks. Coaxial cables are mainly used for access networks. Recently, fiber to the home (FTTH) is being widely deployed. For cable TV companies to utilize readily deployed FTTH, legacy coax network-based customer premise equipment (CPE) such as cable modems and set top boxes may have to be replaced with fiber network based CPE. In order to reduce the expenditure incurred in this process, radio over Internet Protocol (RoIP) technology has been proposed. In the RoIP system, RF signal for coax generated by legacy CPE toward the headend is first sampled, and digitized, and then encapsulated in IP packets, which are then sent to the headend using FTTH access technologies such as passive optical network (PON). In the process of digitizing RF signal, a huge amount of data much more than the base band data is generated. So, a data compression method is required to lower the upstream bit rate maintaining the signal quality. Most compression methods are significantly degraded due to the interference caused by burst noise in the upstream link. Therefore, this paper proposes a new compression method which combines linear predictive coding (LPC), block scaling and non-linear quantization for CATV upstream links. The proposed method is demonstrated to greatly reduce the impact of burst noise while maintaining the compression performance.

Diagonal entropy and topological phase transitions in extended Kitaev chains

We investigate the diagonal entropy for ground states of the extended Kitaev chains with extensive pairing and hopping terms. The systems contain rich topological phases equivalently represented by topological invariant winding numbers and Majorana zero modes. Both the finite size scaling law and block scaling law of the diagonal entropy are studied, which indicates that the diagonal entropy demonstrates volume effect. The parameter of volume term is regarded as the diagonal entropy density, which can identify the critical points of symmetry-protected topological phase transitions efficiently in the studied models, even for those with higher winding numbers. The formulation of block scaling law and the capability of diagonal entropy density in detecting topological phase transitions are independent of the chosen bases. In order to manifest the advantage of diagonal entropy, we also calculate the global entanglement, which cannot show clear signatures of the topological phase transitions. This work provides a new quantum-informatic approach to characterize the feature of the topologically ordered states and may motivate a deep understanding of the quantum coherence and diagonal entropy in various condensed matter systems.

Semadaas Proof of Stake Protocol

The Secure Proof of Stake protocol (SPoS) uses a reputation-verification platform to solve the centralization, efficiency, and security problems that afflict existing blockchain consensus protocols. At heart, a secure proof of stake protocol follows naturally from a secure decentralized autonomous system for validating the reputation of anonymous users. SPoS has several advantageous features over previous PoS protocols. SPoS enables on-chain governance to decide issues such as standards for block creation, policing, and scaling. All protocol upgrades are handled swiftly, decisively, openly, and stably, through the weighted democracy of reputation-staked decision making. Because nodes no longer hash mine but rather check the validity of blocks and vote, SPoS is more efficient than proof-of-work consensus. This paper explains the SPoS protocol and how it maintains decentralization, efficiency, liveness, swift finality, and anonymity, while addressing the security issues of Byzantine faults, the various 51% attacks, stake grinding, the tragedy of the commons, DoS attacks, censorship, liveness faults, Sybil attacks, tyranny of the majority, and soft protocol upgrades.

Advanced predictive methods for wine age prediction: Part II – A comparison study of multiblock regression approaches

In this article, we extend the scope of the first paper of the sequel, which was dedicated to the analysis of advanced single-block regression methods (Rendall et al., 2016) [1], to the class of multiblock regression approaches. The datasets contemplated for developing the multiblock approaches are the same as in the former publication: volatile, polyphenols, organic acids composition and the UV–Vis spectra. The context is still the prediction of the ageing time of one of finest Portuguese fortified wines, the Madeira Wine, but now the data collected from the different analytical sources is explored simultaneously, in a more structured and informative way, through multiblock methodologies. The goal of this paper is to provide a critical assessment of a rich variety of multiblock regression methods, namely: Concatenated PLS, Multiblock PLS (MBPLS), Hierarchical PLS (HPLS), Network-Induced Supervised Learning (NI-SL) and Sequential Orthogonalised Partial Least Squares (SO-PLS). A comparison of block scaling methods was also undertaken for the Concatenated PLS algorithm, and new block scaling methods were proposed that led to better prediction performances. This study explores and reveals the potential advantages of applying multiblock methods for fusing datasets from different sources, from both the predictive and interpretability perspectives.

Scaling of quantum discord in spin models

We study the scaling of quantum discord (a measure of quantum correlation beyond entanglement) in spin models analytically and systematically. We find that at finite temperature the block scaling of quantum discord satisfies an area law for any two-local Hamiltonian. We show that generically and heuristically the two-site scaling of quantum discord is similar to that of correlation functions. In particular, at zero temperature it decays exponentially and polynomially in gapped and gapless (critical) systems, respectively; at finite temperature it decays exponentially in both gapped and gapless systems. We compute the two-site scaling of quantum discord in the XXZ chain, the XY chain (in a magnetic field), and the transverse field Ising chain at zero temperature.

LTE/LTE-A signal compression on the CPRI interface

The Centralized, Cooperative, Cloud Radio Access Network (C-RAN) is a next-generation wireless access network architecture based on centralized processing, collaborative radio, and real time cloud infrastructure. In this architecture, different access technologies (Global System for Mobile Communications (GSM)/Time Division Synchronous Code Division Multiple Access (TD-SCDMA)/Wideband Code Division Multiple Access (WCDMA)/Long Term Evolution (LTE)) can be supported on the same hardware platform in a baseband pool system, which can largely reduce system costs. Long Term Evolution (LTE) and Long Term Evolution-Advanced (LTE-A), which are based on Orthogonal Frequency Division Multiplexing (OFDM) and multiple input multiple output (MIMO) technologies, are regarded as the main wireless access technologies in the evolution from 3G to 4G. A variety of novel technologies such as multi-antenna MIMO, carrier aggregation (CA), and coordinated multipoint have been introduced in LTE/LTE-A to improve system performance, especially in the C-RAN architecture. However, one of the technical challenges for the C-RAN architecture is the fiber bandwidth required for data transmission between the remote radio unit (RRU) and the baseband unit (BBU). We propose using a low-latency baseband signal compression algorithm to solve this problem by reducing the fiber data rate. Using the characteristics of the LTE signal data, we remove the redundancy in the spectral domain. We also leverage block scaling in conjunction with using a linear or nonlinear (non-uniform) quantizer to minimize quantization error. This algorithm effectively reduces the amount of data transmitted between the BBU and RRU, and facilitates the deployment of LTE in the C-RAN architecture. We verified the robustness of the algorithm via simulations and lab tests. The proposed algorithm yields good system performance at a 1/2 compression rate and at a 1/3 compression rate in a practical propagation environment.

Stretchability-aware block scaling for image retargeting

This paper proposes an efficient approach to retarget images based on stretchability-aware block scaling. The image stretchability is first evaluated based on gradient, saliency and color features, and is used to generate the stretchable space. Then the optimal size of the stretched image is determined under the constraint of stretchable space and the same aspect ratio as the target image. Based on the analysis of image stretchability measures, the original image is partitioned into non-stretchable blocks and stretchable blocks, and their scaling factors are calculated based on their stretchability measures and the stretched image size, in order to possibly preserve non-stretchable blocks without distortion and reasonably resize stretchable blocks. Finally, the stretched image is uniformly scaled to generate the target image. Experimental results on a variety of images and the user study demonstrate that our approach achieves an overall better retargeting performance compared to the state-of-the-art image retargeting approaches.

Block scaling in the directed percolation universality class

The universal behaviour of the directed percolation universality class is well understood—both the critical scaling and the finite size scaling. This paper focuses on the block (finite size) scaling of the order parameter and its fluctuations, considering (sub-)blocks of linear size l in systems of linear size L. The scaling depends on the choice of the ensemble, as only the conditional ensemble produces the block-scaling behaviour as established in equilibrium critical phenomena. The dependence on the ensemble can be understood by an additional symmetry present in the unconditional ensemble. The unconventional scaling found in the unconditional ensemble is a reminder of the possibility that scaling functions themselves have a power-law dependence on their arguments.

Backward Error of Polynomial Eigenproblems Solved by Linearization

The most widely used approach for solving the polynomial eigenvalue problem $P(\lambda)x = (\sum_{i=0}^m \l^i A_i) x = 0$ in $n\times n$ matrices $A_i$ is to linearize to produce a larger order pencil $L(\lambda) = \lambda X + Y$, whose eigensystem is then found by any method for generalized eigenproblems. For a given polynomial P, infinitely many linearizations L exist and approximate eigenpairs of P computed via linearization can have widely varying backward errors. We show that if a certain one-sided factorization relating L to P can be found then a simple formula permits recovery of right eigenvectors of P from those of L, and the backward error of an approximate eigenpair of P can be bounded in terms of the backward error for the corresponding approximate eigenpair of L. A similar factorization has the same implications for left eigenvectors. We use this technique to derive backward error bounds depending only on the norms of the $A_i$ for the companion pencils and for the vector space $\mathbb{DL}(P)$ of pencils recently identified by Mackey, Mackey, Mehl, and Mehrmann. In all cases, sufficient conditions are identified for an optimal backward error for P. These results are shown to be entirely consistent with those of Higham, Mackey, and Tisseur on the conditioning of linearizations of P. Other contributions of this work are a block scaling of the companion pencils that yields improved backward error bounds; a demonstration that the bounds are applicable to certain structured linearizations of structured polynomials; and backward error bounds specialized to the quadratic case, including analysis of the benefits of a scaling recently proposed by Fan, Lin, and Van Dooren. The results herein make no assumptions on the stability of the method applied to L or whether the method is direct or iterative.

Two Pass Improved Encoding and its Parallel Processing for Fractal Image Compression

An improvement scheme, so named the Two-Pass Improved Encoding Scheme (TIES), for the application to image compression through the extension of the existing concept of fractal image compression (FIC), which capitalizes on the self-similarity within a given image which is to be compressed, is proposed in this paper. This paper first briefly explores the existing image compression technology based on FIC, before exploring the areas which can be improved and hence establishing the concept behind the TIES algorithm. An effective encoding and decoding algorithm for the implementation of TIES is developed, through the consideration of the domain pool, block scaling and transformation, range block approximation using linear combinations and arithmetic encoding for storing data as close to source entropy as possible. The performance of TIES is then explicitly compared against that of FIC under the same conditions. Finally, due to the long encoding time required by TIES, this paper then proceeds to propose parallelized versions of the two TIES algorithms, before finally concluding with an empirical analysis of the speedup and scalability of the parallelized TIES algorithms, as well as compare the effect of parallelization between the two.

Improved asymptotic stability analysis for Lur'e systems with sector and slope restricted nonlinearities

In this Letter, we propose new criteria for asymptotic stability of Lur'e systems with sector and slope restrictions. The criteria are expressed in terms of linear matrix inequality which has full block scaling matrix. By interpreting the criteria in the frequency domain using KYP lemma, the less conservativeness of the proposed analysis method is investigated. Furthermore, we analyze the dynamic system with input saturation using sector and slop properties of the saturation nonlinearity. A numerical example shows the effectiveness of the proposed method.

Evaluation of different techniques for data fusion of LC/MS and 1H-NMR

In the analyses of highly complex samples (for example, metabolic fingerprinting), the data might not suffice for classification when using only a single analytical technique. Hence, the use of two complementary techniques, e.g., LC/MS and 1H-NMR, might be advantageous. Another possible advantage from using two different techniques is the ability to verify the results (for instance, by verifying a time trend of a metabolic pattern). In this work, both LC/MS and 1H-NMR data from analysis of rat urine have been used to obtain metabolic fingerprints. A comparison of three different methods for data fusion of the two data sets was performed and the possibilities and difficulties associated with data fusion were discussed. When comparing concatenated data, full hierarchical modeling, and batch modeling, the first two approaches were found to be the most successful. Different types of block scaling and variable scaling were evaluated and the optimal scaling for each case was found by cross validation. Validations of the final models were performed by means of an external test set. 2 2 Copies of the Matlab program files used for this work are available from the authors.

Two-stage motion compensation using adaptive global MC and local affine MC

This paper describes a high-efficiency video coding method based on ITU-T H.263. To improve the coding efficiency of H.263, a two-stage motion compensation (MC) method is proposed, consisting of global MC (GMC) for predicting camera motion and local MC (LMC) for macroblock prediction. First, global motion such as panning, tilting, and zooming is estimated, and the global-motion-compensated image is produced for use as a reference in LMC. Next, LMC is performed both for the global-motion-compensated reference image and for the image without GMC. LMC employs an affine motion model in the context of H.263's overlapped block motion compensation. Using the overlapped block affine MC, rotation and scaling of small objects can be predicted, in addition to translational motion. In the proposed method, GMC is adaptively turned on/off for each macroblock since GMC cannot be used for prediction in all regions in a frame. In addition, either an affine or a translational motion model is adaptively selected in LMC for each macroblock. Simulation results show that the proposed video coding technique using the two-stage MC significantly outperforms H.263 under identical conditions, especially for sequences with fast camera motion. The performance improvements in peak-to-peak SNR (PSNR) are about 3 dB over the original H.263, which does not use the two-stage MC.

An efficient bit-serial FIR filter architecture

A new bit-serial architecture for implementation of high order FIR filters is introduced, as well as example FPGA and CMOS realizations. This structure exploits the simplicity of coefficients that consist of two power-of-two terms to yield efficient implementations. Quantization effects are discussed and a simple block scaling method for reducing rounding and truncation noise in high order filters is also presented.