Computational Units Research Articles

Experimental demonstration of an on-chip p-bit core based on stochastic magnetic tunnel junctions and 2D MoS2 transistors

Probabilistic computing is a computing scheme that offers a more efficient approach than conventional complementary metal-oxide–semiconductor (CMOS)-based logic in a variety of applications ranging from optimization to Bayesian inference, and invertible Boolean logic. The probabilistic bit (or p-bit, the base unit of probabilistic computing) is a naturally fluctuating entity that requires tunable stochasticity; by coupling low-barrier stochastic magnetic tunnel junctions (MTJs) with a transistor circuit, a compact implementation is achieved. In this work, by combining stochastic MTJs with 2D-MoS2 field-effect transistors (FETs), we demonstrate an on-chip realization of a p-bit building block displaying voltage-controllable stochasticity. Supported by circuit simulations, we analyze the three transistor-one magnetic tunnel junction (3T-1MTJ) p-bit design, evaluating how the characteristics of each component influence the overall p-bit output. While the current approach has not reached the level of maturity required to compete with CMOS-compatible MTJ technology, the design rules presented in this work are valuable for future experimental implementations of scaled on-chip p-bit networks with reduced footprint.

A GPU based accelerated solver for simulation of heat transfer during metal casting process

The metal casting process, which is one of the key drivers of the manufacturing industry, involves several physical phenomena occurring simultaneously like fluid flow, phase change, and heat transfer which affect the casting yield and quality. Casting process modeling involves numerical modeling of these phenomena on a computer. In recent decades, this has become an inevitable tool for foundry engineers to make defect-free castings. To expedite computational time graphics processing units (GPUs) are being increasingly used in the numerical modeling of heat transfer and fluid flow. Initially, in this work a CPU based implicit solver code is developed for solving the 3D unsteady energy equation including phase change numerically using finite volume method which predicts the thermal profile during solidification in the metal casting process in a completely filled mold. To address the computational bottleneck, which is identified as the linear algebraic solver based on the bi-conjugate gradient stabilized method, a GPU-based code is developed using Compute Unified Device Architecture toolkit and was implemented on the GPU. The CPU and GPU based codes are then validated against a commercial casting simulation code FLOW-3D CAST® for a simple casting part and against in-house experimental results for gravity die casting of a simple geometry. Parallel performance is analyzed for grid sizes ranging from 10 × 10 × 10 to 210 × 210 × 210 and for three time-step sizes. The performance of the GPU code based on occupancy and throughput is also investigated. The GPU code exhibits a maximum speedup of 308× compared to the CPU code for a grid size of 210 × 210 × 210 and a time-step size of 2 s.

A binary neural computing unit with programmable gate using SFQ and CMOS hybrid circuit

Superconducting neural networks hold significant potential for future applications such as natural language processing and image recognition. To this end, we propose a binary neural computing unit implemented using a hybrid circuit of cryogenic CMOS and superconducting technologies. It offers two main advantages: firstly, we utilize current-mode computations for neural unit weight calculations, significantly reducing the unit’s footprint and enabling the potential for higher integration in the future. Secondly, all computations are performed in a low-temperature environment, which implies the possibility of on-chip learning in superconducting neural networks and the potential for achieving faster training rates in the future. We fabricated the chip using Nb 1 kA cm−2 process (1KP) technology and experimentally verified the correctness of the circuit logic. The margins for various control parameters of the circuit are approximately around 30%, and the superconducting circuit power consumption is estimated to be around 4 microwatts.

Abstract PO4-26-01: Computational pathology: Revolutionizing diagnostics and clearing the way for precision medicine

Abstract While targeted cancer therapies often rely on subjective and semi-quantitative visual assessment of protein biomarkers by pathologists through immunohistochemically stained tissue, the transformative force of computational pathology is reshaping healthcare by unleashing unprecedented diagnostic accuracy and unlocking personalized treatments. In recent years, we have established a large integrated computational pathology unit to foster collaboration between interdisciplinary teams of computer scientists, pathologists, molecular biologists, and data scientists. This integration of cutting-edge technologies enabled us to develop a computational pathology approach called Quantitative Continuous Scoring (QCS). QCS deploys the power of Deep Learning (DL) to provide objective and continuous expression data of biomarkers in digitized IHC whole slide images (WSI), particularly of proteins expressed at low levels. While manual scoring of IHC WSIs is limited by subjectivity and semi-quantitative assessment of protein expression, QCS overcomes these limitations with an unprecedented accuracy. QCS utilizes two DL-based algorithms, which we developed fully supervised by using pathologist input as the reference standard. These algorithms identify invasive tumour areas and segment each tumour cell across the WSI into cell nuclei, cytoplasm and membrane. Based on an accurate subcellular segmentation, we can compute biomarker expression, on a continuous scale, as mean Optical Density (OD) in each subcellular compartment based on the Hue-Saturation-Density (HSD) model. Therefore, this approach enables precise detection of the low biomarker expression range with single-cell resolution. Importantly, it also allows the computation of the spatial distribution of tumour cells across the WSI. We have successfully used QCS to drive the selection of antibody clones for IHC assays and to delineate the mode of action and PK/PD mechanisms. Of note, the combination of assessing continuous target expression and capturing the spatial distribution of tumor cells has provided surrogate markers to predict potential bystander activity of antibody drug conjugates (ADCs). This approach outperformed traditional pathologist scoring in identifying patient populations having maximum treatment benefit through retrospective analysis of multiple clinical trials. At present, all computational pathology approaches are developed based on conventional IHC assays that have been optimized for manual scoring. Importantly, we draw a vision in which the assay serves as a critical catalyst for unleashing the full potential of computational pathology by providing high-quality, standardized data inputs. We suggest an approach utilizing orthogonal methods as a reference standard to develop highly sensitive IHC assays, capable of detecting even subtle molecular and cellular changes with precision and exhibiting exceptional specificity for accurately identifying and distinguishing target biomarkers from background noise. In summary, we here describe and discuss a computational pathology-based approach for precise biomarker quantification and superior patient selection with broad applicability and the potential to transform the very fabric of how we diagnose and treat cancer. Citation Format: Mark Gustavson, Markus Schick, Ansh Kapil, Anatoliy Shumilov, Carl Barrett, Hadassah Sade. Computational pathology: Revolutionizing diagnostics and clearing the way for precision medicine [abstract]. In: Proceedings of the 2023 San Antonio Breast Cancer Symposium; 2023 Dec 5-9; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2024;84(9 Suppl):Abstract nr PO4-26-01.

Representation-Learning-Based CNN for Intelligent Attack Localization and Recovery of Cyber-Physical Power Systems.

Enabled by the advances in communication networks, computational units, and control systems, cyber-physical power systems (CPPSs) are anticipated to be complex and smart systems in which a large amount of data are generated, exchanged, and processed for various purposes. Due to these strong interactions, CPPSs will introduce new security vulnerabilities. To ensure secure operation and control of CPPSs, it is essential to detect the locations of the attacked measurements and remove the state bias caused by malicious cyber-attacks such as false data inject attack, jamming attack, denial of service attack, or hybrid attack. Accordingly, this article makes the first contribution concerning the representation-learning-based convolutional neural network (RL-CNN) for intelligent attack localization and system recovery of CPPSs. In the proposed method, the cyber-attacks' locational detection problem is formulated as a multilabel classification problem for CPPSs. An RL-CNN is originally adopted as the multilabel classifier to explore and exploit the implicit information of measurements. By comparing with previous multilabel classifiers, the RL-CNN improves the performance of attack localization for complex CPPSs. Then, to automatically filter out the cyber-attacks for system recovery, a mean-squared estimator is used to handle the difficulty in state estimation with the removal of contaminated measurements. In this scheme, prior knowledge of the system state is obtained based on the outputs of the stochastic power flow or historical measurements. The extensive simulation results in three IEEE bus systems show that the proposed method is able to provide high accuracy for attack localization and perform automatic attack filtering for system recovery under various cyber-attacks.

Wflow_sbm v0.7.3, a spatially distributed hydrological model: from global data to local applications

Abstract. The wflow_sbm hydrological model, recently released by Deltares, as part of the Wflow.jl (v0.7.3) modelling framework, is being used to better understand and potentially address multiple operational and water resource planning challenges from a catchment scale to national scale to continental and global scale. Wflow.jl is a free and open-source distributed hydrological modelling framework written in the Julia programming language. The development of wflow_sbm, the model structure, equations and functionalities are described in detail, including example applications of wflow_sbm. The wflow_sbm model aims to strike a balance between low-resolution, low-complexity and high-resolution, high-complexity hydrological models. Most wflow_sbm parameters are based on physical characteristics or processes, and at the same time wflow_sbm has a runtime performance well suited for large-scale high-resolution model applications. Wflow_sbm models can be set a priori for any catchment with the Python tool HydroMT-Wflow based on globally available datasets and through the use of point-scale (pedo)transfer functions and suitable upscaling rules and generally result in a satisfactory (0.4 ≥ Kling–Gupta efficiency (KGE) < 0.7) to good (KGE ≥ 0.7) performance for discharge a priori (without further tuning). Wflow_sbm includes relevant hydrological processes such as glacier and snow processes, evapotranspiration processes, unsaturated zone dynamics, (shallow) groundwater, and surface flow routing including lakes and reservoirs. Further planned developments include improvements on the computational efficiency and flexibility of the routing scheme, implementation of a water demand and allocation module for water resource modelling, the addition of a deep groundwater concept, and computational efficiency improvements through for example distributed computing and graphics processing unit (GPU) acceleration.

Design of floating point multiplier using approximate hybrid Radix-4/ Radix-8 booth encoder for image analysis

This paper presents a novel power efficient Hybrid floating point multiplier (HFPM) with approximate hybrid radix-4/radix-8 booth encoder. This approach efficiently increases the speed of operation and optimises hardware. Compared to existing HFPMs, the proposed HFPM utilises a novel hybrid radix-4/radix-8 booth encoder. The hybrid architecture reduces the computations by an average of k/3 for a k-bit 2’s complement representation by reducing the number of partial products. In addition, to achieve an area-efficient and power-efficient design, approximate computational units are added to the radix-8 booth encoder in the hybrid architecture with a slight reduction in accuracy. The efficiency of the approximate architecture is measured in scales of Error Rate (ER) and Normalized Error Distance (NED). The proposed design is implemented on Kintex 7 FPGA and realized on 45nm ASIC platforms. The experimental study shows that the proposed multiplier has utilised hardware (831 slice LUT, 828 LUT as logic) and consumed 6110 μm2 of area and 83 μW of power which is less as compared to existing floating point multipliers. The experimental results in terms of average Structural Similarity Index Measure (SSIM) is 0.98 which is better than the existing floating point multipliers are compared.

Intelligent Tribotronic Transistors Toward Tactile Near‐Sensor Computing

AbstractFor the next generation of human‐machine interaction (HMI) systems, the development of a tactile interaction unit with multimodal, high sensitivity, and real‐time perception and recognition is the key. Herein, an artificial tactile near‐sensor computing (ATNSC) unit based on a triboelectric tactile sensor and an organic synaptic transistor is reported. By introducing multi‐peak microstructures, the mechanical performance of the tactile sensor is optimized, showing a high sensitivity of 0.98 V kPa−1 in the pressure range of 0–10 kPa and maintaining 0.11 V kPa−1 at high pressures up to 350 kPa. Additionally, by designing stripe‐like convex structures on the top surface, the sensor is capable of bimodal perception in both pressure and sliding sensations. Furthermore, the organic synaptic transistor, which can be driven by tactile sensing stimuli in a variety of circumstances, is achieved utilizing an ion‐rich gelatin dielectric covered by a hydrophobic polymer coating layer. The ATNSC unit well demonstrates the stimuli‐dependent short‐term memory effect, and it enables tactile near‐sensor computing for feature action recognition in an HMI system, laying a solid foundation for the construction of intelligent interaction devices.

Dynamic characteristics of refrigerant evaporation temperature in air-water heat source heat pump

In order to improve the problem of reduced heat capacity and unstable heating ability of air source heat pump (ASHP) in the heating mode during the cold season, a heat pump evaporator using air and surface water as a heat source is designed in this paper. A heat transfer unit's calculation model of an air–water heat source evaporator is established. Experimental tests and simulation studies are conducted on the refrigerant evaporation temperature and heat pump heating performance. On this basis, a refrigerant evaporation temperature prediction model is constructed using a multiple linear regression (MLR) model. The trend of refrigerant evaporation temperature under air–water heat source is analyzed by taking the winter meteorological parameters in Xiangtan City as an example. The results show that the air–water heat source heat pump (A-WSHP) has a higher refrigerant evaporation temperature and a smaller temperature fluctuation range than the ASHP. The average values of refrigerant evaporation temperatures under the experimental conditions were 5.3 °C and 4.6 °C, respectively. The temperature fluctuation ranges were 3.8 °C-6.7 °C and 2.8 °C-7.0 °C, respectively. The A-WSHP has higher heating capacity and heating stability. Air temperature, air velocity, water spray temperature and water spray flow rate are all important parameters that affect the refrigerant evaporation temperature of a heat pump. Evaporation temperature changes are positively correlated with these factors. In the refrigerant evaporation temperature prediction model, the air temperature change has the most significant effect on the evaporation temperature, followed by spray water temperature, air velocity, and spray flow rate, whose influence weights are 82.5 %, 38.8 %, 29.4 %, and 25.3 %, respectively. Selecting different temperatures of spray water and adjusting the evaporator's inlet spray flow rate can improve the refrigerant evaporation temperature and heat pump heating stability.

Centripetal integration of past events in hippocampal astrocytes regulated by locus coeruleus.

An essential feature of neurons is their ability to centrally integrate information from their dendrites. The activity of astrocytes, in contrast, has been described as mostly uncoordinated across cellular compartments without clear central integration. Here we report conditional integration of calcium signals in astrocytic distal processes at their soma. In the hippocampus of adult mice of both sexes, we found that global astrocytic activity, as recorded with population calcium imaging, reflected past neuronal and behavioral events on a timescale of seconds. Salient past events, indicated by pupil dilations, facilitated the propagation of calcium signals from distal processes to the soma. Centripetal propagation to the soma was reproduced by optogenetic activation of the locus coeruleus, a key regulator of arousal, and reduced by pharmacological inhibition of α1-adrenergic receptors. Together, our results suggest that astrocytes are computational units of the brain that slowly and conditionally integrate calcium signals upon behaviorally relevant events.

Topologically Protected Quantum Logic Gates with Valley-Hall Photonic Crystals.

Topological photonics provide a promising way to realize more robust optical devices against some defects and environmental perturbations. Quantum logic gates are fundamental units of quantum computers, which are widely used in future quantum information processing. Thus, constructing robust universal quantum logic gates is an important way forward to practical quantum computing. However, the most important problem to be solved is how to construct the quantum-logic-gate-required 2×2 beam splitter with topological protection. Here, the experimental realization of the topologically protected contradirectional coupler is reported, which can be employed to realize the quantum logic gates, including control-NOT and Hadamard gates, on the silicon photonic platform. These quantum gates not only have high experimental fidelities but also exhibit a certain degree of tolerances against certain types of defects. This work paves the way for the development of practical optical quantum computations and signal processing.

A universal parallel simulation framework for energy pipeline networks on high-performance computers

Energy distribution networks represent crucial infrastructures for modern society, and various simulation tools have been widely used by energy suppliers to manage these intricate networks. However, simulation calculations include a large number of fluid control equations, and computational overhead limits the performance of simulation software. This paper proposes a universal parallel simulation framework for energy pipeline networks that takes advantages of data parallelism and computational independence between network elements. A non-pipe model of an energy supply network is optimized, and the input and output of the network model in the proposed framework are modified, which can reduce the development burden during the numerical computations of the pipeline network and weaken the computational correlation between different simulated components. In addition, independent computations can be performed concurrently through periodic data exchange procedures between component instances, improving the parallelism and efficiency of simulation computations. Further, a parallel water pipelines network simulation computing paradigm based on a heterogeneous computer hardware architecture is used to evaluate the proposed framework’s performance. A series of tests are conducted to verify the accuracy of the proposed framework, and simulation errors of less than 5% are achieved. The results of multi-threaded simulation experiments have demonstrated the feasibility of the proposed framework in a parallel computing approach. Moreover, an Advanced Micro Devices (AMD) Deep Computing Unit (DCU)-parallel program is implemented into a water supply network simulation system; the computational efficiency of this system is compared with that of its serial counterpart. The experimental results show that the proposed framework is appropriate for high-performance computer architectures, and the 18x speed-up ratio demonstrates that the parallel program based on the proposed universal framework outperforms the serial program. That provides the basis for the application of pipe network simulation on high-performance computers.

Dataflow optimization with layer-wise design variables estimation method for enflame CNN accelerators

As convolution layers have been proved to be the most time-consuming operation in convolutional neural network (CNN) algorithms, many efficient CNN accelerators have been designed to boost the performance of convolution operations. Previous works on CNN acceleration usually use fixed design variables for diverse convolutional layers, which would lead to inefficient data movements and low utilization of computing resource. We tackle this issue by proposing a flexible dataflow optimization method with design variables estimation for different layers. The optimization method first narrows the design space by the priori constraints, and then enumerates all legal solutions to select the optimal design variables. We demonstrate the effectiveness of the proposed optimization method by implementing representative CNN models (VGG-16, ResNet-18 and MobileNet V1) on Enflame Technology's programmable CNN accelerator, General Computing Unit (GCU). The results indicate that our optimization can significantly enhance the throughput of the convolution layers in ResNet, VGG and MobileNet on GCU, with improvement of up to 1.84×. Furthermore, it achieves up to 2.08× of GCU utilization specifically for the convolution layers of ResNet on GCU.

Integrated Route-Planning System for Agricultural Robots

Within the transition from precision agriculture (task-specific approach) to smart farming (system-specific approach) there is a need to build and evaluate robotic systems that are part of an overall integrated system under a continuous two-way connection and interaction. This paper presented an initial step in creating an integrated system for agri-robotics, enabling two-way communication between an unmanned ground vehicle (UGV) and a farm management information system (FMIS) under the general scope of smart farming implementation. In this initial step, the primary task of route-planning for the agricultural vehicles, as a prerequisite for the execution of any field operation, was selected as a use-case for building and evaluating this integration. The system that was developed involves advanced route-planning algorithms within the cloud-based FMIS, a comprehensive algorithmic package compatible with agricultural vehicles utilizing the Robot Operating System (ROS), and a communicational and computational unit (CCU) interconnecting the FMIS algorithms, the corresponding user interface, and the vehicles. Its analytical module provides valuable information about UGVs’ performance metrics, specifically performance indicators of working distance, non-working distance, overlapped area, and field-traversing efficiency. The system was demonstrated via the implementation of two robotic vehicles in route-execution tasks in various operational configurations, field features, and cropping systems (open field, row crops, orchards). The case studies showed variability in the operational performance of the field traversal efficiency to be between 79.2% and 93%, while, when implementing the optimal route-planning functionality of the system, there was an improvement of up to 9.5% in the field efficiency. The demonstrated results indicate that the user can obtain better control over field operations by making alterations to ensure optimum field performance, and the user can have complete supervision of the operation.

One-vs-One, One-vs-Rest, and a novel Outcome-Driven One-vs-One binary classifiers enabled by optoelectronic memristors towards overcoming hardware limitations in multiclass classification

Deep neural networks have achieved considerable success over the past ten years in a variety of fields. However, current state-of-the-art artificial intelligence (AI) systems require large computing hardware infrastructure and high power consumption. To overcome these hurdles, it is required to adopt new strategies such as designing novel computation architectures and developing building blocks that can mimic the low energy consumption of biological systems. On the architecture level, implementing classification tasks by splitting the problem into simpler subtasks is a way to relax hardware constraints despite the less accuracy of the approach. On the computation unit level, memristive devices are a promising technology for low power neuromorphic computation. Hereby, we combine both these two approaches and present a novel algorithmic approach for multiclass classification tasks through splitting the problem into binary subtasks while using optoelectronics memristors as synapses. Our approach leverages the core fundamentals from the One-vs-One (OvO) and the One-vs-Rest (OvR) classification strategies towards a novel Outcome-Driven One-vs-One (ODOvO) approach. The light modulation of synaptic weights, fed in our algorithm from experimental data, is a key enabling parameter that permits classification without modifying further applied electrical biases. Our approach requires at least a 10X less synapses (only 196 synapses are required) while reduces the classification time by up to N2\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\frac{N}{2}$$\\end{document} compared to conventional memristors. We show that the novel ODOvO algorithm has similar accuracies to OvO (reaching over 60% on the MNIST dataset) while requiring even fewer iterations compared to the OvR. Consequently, our approach constitutes a feasible solution for neural networks where key priorities are the minimum energy consumption i.e., small iterations number, fast execution, and the low hardware requirements allowing experimental verification.

Distributed feature representations of natural stimuli across parallel retinal pathways

How sensory systems extract salient features from natural environments and organize them across neural pathways is unclear. Combining single-cell and population two-photon calcium imaging in mice, we discover that retinal ON bipolar cells (second-order neurons of the visual system) are divided into two blocks of four types. The two blocks distribute temporal and spatial information encoding, respectively. ON bipolar cell axons co-stratify within each block, but separate laminarly between them (upper block: diverse temporal, uniform spatial tuning; lower block: diverse spatial, uniform temporal tuning). ON bipolar cells extract temporal and spatial features similarly from artificial and naturalistic stimuli. In addition, they differ in sensitivity to coherent motion in naturalistic movies. Motion information is distributed across ON bipolar cells in the upper and the lower blocks, multiplexed with temporal and spatial contrast, independent features of natural scenes. Comparing the responses of different boutons within the same arbor, we find that axons of all ON bipolar cell types function as computational units. Thus, our results provide insights into the visual feature extraction from naturalistic stimuli and reveal how structural and functional organization cooperate to generate parallel ON pathways for temporal and spatial information in the mammalian retina.

Transcutaneous Magnet Localizer for a Self-Contained Myokinetic Prosthetic Hand.

The search for a physiologically appropriate interface for the control of dexterous hand prostheses is an ongoing challenge in bioengineering. In this context, we proposed an interface, named myokinetic control interface, based on the localization of magnets implanted in the residual limb muscles, to monitor their contractions and send appropriate commands to the artificial hand. As part of such concept, this interface requires a transcutaneous magnet localizer that can be integrated in a self-contained limb prosthesis, a feature yet to be realized within the current state of the art. In an attempt to cover this gap, here we present a modular embedded system consisting of a computation unit able to acquire synchronized samples captured by up to eight acquisition units, so to localize multiple magnets. The system exhibits short computation times (<60ms) and power consumption (0.6-1.2W) which are suitable for use in a clinically viable prosthetic arm. The system proved able to localize magnets while moving at speeds in the range of physiological movements (<0.24m/s), with high accuracy (<1mm) and precision (<0.5mm). We demonstrated a system suitable for the implementation of a self-contained myokinetic prosthetic hand. These results pave the way towards the clinical implementation of the myokinetic interface, with amputees controlling an artificial arm by means of implanted magnets.

Biophysical Modeling of Synaptic Plasticity.

Dendritic spines are small, bulbous compartments that function as postsynaptic sites and undergo intense biochemical and biophysical activity. The role of the myriad signaling pathways that are implicated in synaptic plasticity is well studied. A recent abundance of quantitative experimental data has made the events associated with synaptic plasticity amenable to quantitative biophysical modeling. Spines are also fascinating biophysical computational units because spine geometry, signal transduction, and mechanics work in a complex feedback loop to tune synaptic plasticity. In this sense, ideas from modeling cell motility can inspire us to develop multiscale approaches for predictive modeling of synaptic plasticity. In this article, we review the key steps in postsynaptic plasticity with a specific focus on the impact of spine geometry on signaling, cytoskeleton rearrangement, and membrane mechanics. We summarize the main experimental observations and highlight how theory and computation can aid our understanding of these complex processes.

Distributed Feature Selection for Power System Dynamic Security Region Based on Grid-Partition and Fuzzy-Rough Sets

In order to satisfy the requirements of modern online security assessment of power systems with continuously increasing complexity in terms of structure and scale, it is desirable to develop a power system dynamic security region (DSR) analysis. However, data-driven methods suffer from expensive model training costs and overfitting when determining DSR boundaries with high-dimensional grid features. Given this problem, a distributed feature selection method based on grid partition and fuzzy-rough sets is proposed in this paper. The method first employs the Louvain algorithm to partition the power grid and divide the original feature set so that high-dimensional features can be allocated to multiple computational units for distributed screening. At this point, the connections between features of different computational units are minimized to a relatively low level, thereby avoiding large errors in the distributed results. Then, an incremental search algorithm based on the fuzzy-rough set theory (FRST) is used for feature selection at each computational unit, which can effectively take into account the intrinsic connections between features. Finally, the results of all computational units are integrated in the coordination unit to complete the overall feature selection. The experimental results based on the IEEE-39 bus system show that the proposed method can help simplify the power system DSR analysis with high-dimensional features by screening the critical features. And compared with other commonly used filter methods, it has higher screening accuracy and lower time costs.

Novel adaptive quantization methodology for 8-bit floating-point DNN training

There is a high energy cost associated with training Deep Neural Networks (DNNs). Off-chip memory access contributes a major portion to the overall energy consumption. Reduction in the number of off-chip memory transactions can be achieved by quantizing the data words to low data bit-width (E.g., 8-bit). However, low-bit-width data formats suffer from a limited dynamic range, resulting in reduced accuracy. In this paper, a novel 8-bit Floating Point (FP8) data format quantized DNN training methodology is presented, which adapts to the required dynamic range on-the-fly. Our methodology relies on varying the bias values of FP8 format to fit the dynamic range to the required range of DNN parameters and input feature maps. The range fitting during the training is adaptively performed by an online statistical analysis hardware unit without stalling the computation units or its data accesses. Our approach is compatible with any DNN compute cores without any major modifications to the architecture. We propose to integrate the new FP8 quantization unit in the memory controller. The FP32 data from the compute core are converted to FP8 in the memory controller before writing to the DRAM and converted back after reading the data from DRAM. Our results show that the DRAM access energy is reduced by 3.07×\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ imes $$\\end{document} while using an 8-bit data format instead of using 32-bit. The accuracy loss of the proposed methodology with 8-bit quantized training is ≈1%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\approx 1\\%$$\\end{document} for various networks with image and natural language processing datasets.