Computational Units Research Articles

Flexible silicon photonic architecture for accelerating distributed deep learning

The increasing size and complexity of deep learning (DL) models have led to the wide adoption of distributed training methods in datacenters (DCs) and high-performance computing (HPC) systems. However, communication among distributed computing units (CUs) has emerged as a major bottleneck in the training process. In this study, we propose Flex-SiPAC, a flexible silicon photonic accelerated compute cluster designed to accelerate multi-tenant distributed DL training workloads. Flex-SiPAC takes a co-design approach that combines a silicon photonic hardware platform with a tailored collective algorithm, optimized to leverage the unique physical properties of the architecture. The hardware platform integrates a novel wavelength-reconfigurable transceiver design and a micro-resonator-based wavelength-reconfigurable switch, enabling the system to achieve flexible bandwidth steering in the wavelength domain. The collective algorithm is designed to support reconfigurable topologies, enabling efficient all-reduce communications that are commonly used in DL training. The feasibility of the Flex-SiPAC architecture is demonstrated through two testbed experiments. First, an optical testbed experiment demonstrates the flexible routing of wavelengths by shuffling an array of input wavelengths using a custom-designed spatial-wavelength selective switch. Second, a four-GPU testbed running two DL workloads shows a 23% improvement in job completion time compared to a similarly sized leaf-spine topology. We further evaluate Flex-SiPAC using large-scale simulations, which show that Flex-SiPAC is able to reduce the communication time by 26% to 29% compared to state-of-the-art compute clusters under representative collective operations.

Optimizing for In-Memory Deep Learning With Emerging Memory Technology.

In-memory deep learning executes neural network models where they are stored, thus avoiding long-distance communication between memory and computation units, resulting in considerable savings in energy and time. In-memory deep learning has already demonstrated orders of magnitude higher performance density and energy efficiency. The use of emerging memory technology (EMT) promises to increase density, energy, and performance even further. However, EMT is intrinsically unstable, resulting in random data read fluctuations. This can translate to nonnegligible accuracy loss, potentially nullifying the gains. In this article, we propose three optimization techniques that can mathematically overcome the instability problem of EMT. They can improve the accuracy of the in-memory deep learning model while maximizing its energy efficiency. Experiments show that our solution can fully recover most models' state-of-the-art (SOTA) accuracy and achieves at least an order of magnitude higher energy efficiency than the SOTA.

Novel Nano-Scale Computing Unit for the IoBNT: Concept and Practical Considerations

Novel Nano-Scale Computing Unit for the IoBNT: Concept and Practical Considerations

PipePIM: Maximizing Computing Unit Utilization in ML-Oriented Digital PIM by Pipelining and Dual Buffering

PipePIM: Maximizing Computing Unit Utilization in ML-Oriented Digital PIM by Pipelining and Dual Buffering

TLS Fast Orbit Feedback Upgrade

The Orbit Feedback System (OFB) at the Taiwan Light Source (TLS) was implemented two decades ago and has undergone multiple upgrades to enhance its performance. The loop bandwidth was constrained by the available hardware. The system is unable to eliminate perturbations from a fast source. In order to enhance the performance of orbit feedback, the system underwent an upgrade in 2008. The upgrade involved the replacement of the analog-type BPM electronics with digital BPM electronics, as well as the replacement of the corrector power supply with a high-performance switching-type power supply that has a wide bandwidth. Following the commissioning of the Taiwan Photon Source (TPS) in 2015, a decision was made to gradually migrate the control system of the Taiwan Light Source to the Experimental Physics and Industrial Control System (EPICS). This decision was made in order to facilitate resource sharing between the two facilities since EPICS had already been adopted by TPS. The Orbit Feedback System is a revitalized subsystem that now benefits from EPICS support. Additionally, the feedback computation unit has been upgraded to a Field Programmable Gate Array (FPGA), resulting in an increase in the calculating cycle from 2.5 kHz to 10 kHz. This report presents a concise overview of the integration of fast orbit feedback and the outcomes of the orbit stability enhancement following the upgrade.

Organization of a Regional Computer Unit for Statistical Econometrics applications, Database Organization Access to International Bibliography

Organization of a Regional Computer Unit for Statistical Econometrics applications, Database Organization Access to International Bibliography

Development of ocular dominance columns across rodents and other species: revisiting the concept of critical period plasticity.

The existence of cortical columns, regarded as computational units underlying both lower and higher-order information processing, has long been associated with highly evolved brains, and previous studies suggested their absence in rodents. However, recent discoveries have unveiled the presence of ocular dominance columns (ODCs) in the primary visual cortex (V1) of Long-Evans rats. These domains exhibit continuity from layer 2 through layer 6, confirming their identity as genuine ODCs. Notably, ODCs are also observed in Brown Norway rats, a strain closely related to wild rats, suggesting the physiological relevance of ODCs in natural survival contexts, although they are lacking in albino rats. This discovery has enabled researchers to explore the development and plasticity of cortical columns using a multidisciplinary approach, leveraging studies involving hundreds of individuals-an endeavor challenging in carnivore and primate species. Notably, developmental trajectories differ depending on the aspect under examination: while the distribution of geniculo-cortical afferent terminals indicates matured ODCs even before eye-opening, consistent with prevailing theories in carnivore/primate studies, examination of cortical neuron spiking activities reveals immature ODCs until postnatal day 35, suggesting delayed maturation of functional synapses which is dependent on visual experience. This developmental gap might be recognized as 'critical period' for ocular dominance plasticity in previous studies. In this article, I summarize cross-species differences in ODCs and geniculo-cortical network, followed by a discussion on the development, plasticity, and evolutionary significance of rat ODCs. I discuss classical and recent studies on critical period plasticity in the venue where critical period plasticity might be a component of experience-dependent development. Consequently, this series of studies prompts a paradigm shift in our understanding of species conservation of cortical columns and the nature of plasticity during the classical critical period.

FPGA based Flexible Implementation of Light Weight Inference on Deep Convolutional Neural Networks

Standard Convolution (StdConv) is the main technique used in the state of the art Deep Convolutional Neural Networks (DCNNs). Fewer computations are achieved if Depthwise Separable Convolution technique (SepConv) is used as an alternative. A crucial issue in many applications like smart cameras and autonomous vehicles where low latency is essential stems from deploying a lightweight and low cost inference models. An acceptable accuracy should be kept with tolerable computations and memory access load. A flexible architecture for different DCNN convolution types and models is proposed. The flexibility comes from the sharing of one memory access unit with different types of layers regardless of the selected kernel size, by multiplying each weight vector by local operators with variant aperture. Moreover, one depthwise computation unit can be used for both standard and pointwise layers. The learnable parameters are quantized to 8-bits fixed point representation and that gives very limited reduction of accuracy and a considerable reduction of the Field-Programmable Gate Array (FPGA) resources. To reduce processing time, inter layer parallel computations are performed. The experiment is conducted by using grey scale ORL database with shallow Convolutional Neural Network (CNN) and the colored Canadian Institute for Advanced Research 10 classes (CIFAR-10) database with DCNN, and a comparable accuracies of 93% and 85.7% are achieved respectively using very low cost of Spartan 3E and moderate cost of zynq FPGA platforms

Универсальный метод Монте-Карло для процессов Леви и его экстремумов

The article proposes a universal approach to constructing Monte Carlo methods for pricing options with payoffs depending on the joint distribution of the final position of the Lévy process X_T and its infimum J_T (or supremum S_T). We derive approximate formulas for the conditional cumulative distribution functions of the Lévy process "P"(X_T<x|S_T=y) (P(X_T<x|J_T=y)), which are expressed through the partial derivative with respect to y of the joint cumulative distribution function "P"(X_T<x,S_T<y) (P(X_T<x,J_T<y)) and the density of the infimum (or supremum) at the final moment of time. By applying the Laplace transform to the joint cumulative distribution function of the Lévy process and its extremum, we use the approximate Wiener–Hopf factorization to represent the image of its partial derivative. By inverting the Laplace transform using the Gaver–Stehfest algorithm, we find the desired conditional cumulative distribution function. The developed algorithm for simulating the joint position of the Lévy process and its extremum at a given point in time consists of two key stages. At the first stage, we simulate the extremum value of the L´evy process based on the approximation of its cumulative distribution function "P"(S_T<x) (or P(J_T<x)). In the second step, we simulate the final value of the Lévy process based on the approximation of the conditional cumulative distribution function of the final position of the Lévy process relative to its extremum. The universality of the Monte Carlo method we developed lies in the implementation of a uniform approach for a wide class of Lévy processes, in contrast to classical approaches, when simulations are essentially based on the features of the probability distribution associated with the simulated random process or its extrema. In our approach, it is enough to know the characteristic exponent of the Lévy process. The most time-consuming computational unit for simulating a random variable based on a known cumulative distribution function can be effectively implemented using neural networks and accelerated through parallel computing. Thus, on the one hand, the approach we propose is suitable for a wide class of Lévy models, on the other hand, it can be combined with machine learning methods.

A Visual Comparison of Silent Error Propagation.

High-performance computing (HPC) systems play a critical role in facilitating scientific discoveries. Their scale and complexity (e.g., the number of computational units and software stack) continue to grow as new systems are expected to process increasingly more data and reduce computing time. However, with more processing elements, the probability that these systems will experience a random bit-flip error that corrupts a program's output also increases, which is often recognized as silent data corruption. Analyzing the resiliency of HPC applications in extreme-scale computing to silent data corruption is crucial but difficult. An HPC application often contains a large number of computation units that need to be tested, and error propagation caused by error corruption is complex and difficult to interpret. To accommodate this challenge, we propose an interactive visualization system that helps HPC researchers understand the resiliency of HPC applications and compare their error propagation. Our system models an application's error propagation to study a program's resiliency by constructing and visualizing its fault tolerance boundary. Coordinating with multiple interactive designs, our system enables domain experts to efficiently explore the complicated spatial and temporal correlation between error propagations. At the end, the system integrated a nonmonotonic error propagation analysis with an adjustable graph propagation visualization to help domain experts examine the details of error propagation and answer such questions as why an error is mitigated or amplified by program execution.

High-performance unified modular multiplication algorithm and hardware architecture over G(2m)

The finite field Modular Multiplier (MM) over GF (2m) based on the National Institute of Standards and Technology (NIST) recommended polynomials can be used as a critical component in many cryptosystems. But much of the research suffers from the problems of only supporting a single curve or being very inefficient. In this paper, we present a KA-based unified algorithm (hybrid field size) and its corresponding architecture for high-performance implementation of the systolic modular multiplier over GF (2m), which can support all the NIST recommended polynomials. A number of efficient techniques have been explored and used to realize a high-performance implementation of the multiplier. First, we propose a three-term KA based MM approach and make it fit for systolic implementation. Second, we propose a configurable, unified irreducible polynomial that can support all five curves recommended by NIST and the corresponding modular reduction method. Third, the corresponding systolic structure is also designed in accordance with (1) and (2) to improve efficiency, and we adopt register sharing and computation unit sharing methods to reduce the space complexity of the proposed structure. Additionally, parallel computation of two163-width MMs is employed to improve resource utilization further. From the synthesis results on ASIC, it is shown that the proposed multipliers have significantly lower latency and higher performance than the existing designs. e.g., the proposed structure could achieve up to a 73 % reduction in area-delay product (ADP) over the existing structures.

Stable coding of aversive associations in medial prefrontal populations.

The medial prefrontal cortex (mPFC) is at the core of numerous psychiatric conditions, including fear and anxiety-related disorders. Whereas an abundance of evidence suggests a crucial role of the mPFC in regulating fear behaviour, the precise role of the mPFC in this process is not yet entirely clear. While studies at the single-cell level have demonstrated the involvement of this area in various aspects of fear processing, such as the encoding of threat-related cues and fear expression, an increasingly prevalent idea in the systems neuroscience field is that populations of neurons are, in fact, the essential unit of computation in many integrative brain regions such as prefrontal areas. What mPFC neuronal populations represent when we face threats? To address this question, we performed electrophysiological single-unit population recordings in the dorsal mPFC while mice faced threat-predicting cues eliciting defensive behaviours, and performed pharmacological and optogenetic inactivations of this area and the amygdala. Our data indicated that the presence of threat-predicting cues induces a stable coding dynamics of internally driven representations in the dorsal mPFC, necessary to drive learned defensive behaviours. Moreover, these neural population representations primary reflect learned associations rather than specific defensive behaviours, and the construct of such representations relies on the functional integrity of the amygdala.

Numerical simulation method and deterioration analysis of load-bearing performance of tunnel linings containing cold joints in plain concrete

Cracking in tunnel linings is a common issue in tunnel engineering, with the main cause being construction-induced defects. As a construction-induced defect in a lining, a cold joint affects the bearing capacity of the lining, and the effect of cold joints on the bearing performance deterioration of linings remains unclear. In this study, 180 high-speed railroad tunnels built in Grade III surrounding rock areas were investigated to determine their degradation states, and the distribution of lining cracks in these tunnels was explored on the basis of different factors. Subsequently, three-point bending experiments were conducted on concrete components with different pouring interval cold joint. Based on the experimental results, a computational unit applicable to cold joint simulation was proposed, and the unit parameters were inverted. Finally, the unit was applied to establish a tunnel model with cold joints and to analyze the deterioration effect of cold joints on the bearing performance of the lining. The cracks caused by cold joints in the lining of a structure were found to extend perpendicular to the lining surface, and this can greatly affect the overall strength of the structure. It's important to consider the impact of cold joints on the lining. In order to simulate these cold joints, cohesive elements were used to calculate the maximum stress and stiffness degradation of the lining with and without cold joint elements. The results showed that there were only small differences, indicating that cohesive elements are effective for simulating cold joints. Two cracking patterns were identified, the dorsal tension misalignment model and the medial tension misalignment model, which are related to the stress state of the cold joints. It's necessary to pay attention to cold joints in the inner tensioned area of the lining and those that exceed the final setting time of concrete during pouring.

Large-scale single-neuron speech sound encoding across the depth of human cortex

Understanding the neural basis of speech perception requires that we study the human brain both at the scale of the fundamental computational unit of neurons and in their organization across the depth of cortex. Here we used high-density Neuropixels arrays1–3 to record from 685 neurons across cortical layers at nine sites in a high-level auditory region that is critical for speech, the superior temporal gyrus4,5, while participants listened to spoken sentences. Single neurons encoded a wide range of speech sound cues, including features of consonants and vowels, relative vocal pitch, onsets, amplitude envelope and sequence statistics. Neurons at each cross-laminar recording exhibited dominant tuning to a primary speech feature while also containing a substantial proportion of neurons that encoded other features contributing to heterogeneous selectivity. Spatially, neurons at similar cortical depths tended to encode similar speech features. Activity across all cortical layers was predictive of high-frequency field potentials (electrocorticography), providing a neuronal origin for macroelectrode recordings from the cortical surface. Together, these results establish single-neuron tuning across the cortical laminae as an important dimension of speech encoding in human superior temporal gyrus.

A General Approach to Dropout in Quantum Neural Networks

AbstractIn classical machine learning (ML), “overfitting” is the phenomenon occurring when a given model learns the training data excessively well, and it thus performs poorly on unseen data. A commonly employed technique in ML is the so called “dropout,” which prevents computational units from becoming too specialized, hence reducing the risk of overfitting. With the advent of quantum neural networks (QNNs) as learning models, overfitting might soon become an issue, owing to the increasing depth of quantum circuits as well as multiple embedding of classical features, which are employed to give the computational nonlinearity. Here, a generalized approach is presented to apply the dropout technique in QNN models, defining and analyzing different quantum dropout strategies to avoid overfitting and achieve a high level of generalization. This study allows to envision the power of quantum dropout in enabling generalization, providing useful guidelines on determining the maximal dropout probability for a given model, based on overparametrization theory. It also highlights how quantum dropout does not impact the features of the QNN models, such as expressibility and entanglement. All these conclusions are supported by extensive numerical simulations and may pave the way to efficiently employing deep quantum machine learning (QML) models based on state‐of‐the‐art QNNs.

Toddler monitoring system in vehicle using single shot detector mobilenet and single shot detector-inception on Jetson Nano

<p>Road vehicles are today’s primary form of transportation; the safety of children passengers must take precedence. Numerous reports of toddler death in road vehicles, include heatstroke and accidents caused by negligent parents. In this research, we report a system developed to monitor and detect a toddler's presence in a vehicle and to classify the toddler's seatbelt status. The objective of the toddler monitoring system is to monitor the child's conditions to ensure the toddler's safety. The device senses the toddler's seatbelt status and warns the driver if the child is left in the car after the vehicle is powered off. The vision-based monitoring system employs deep learning algorithms to recognize infants and seatbelts, in the interior vehicle environment. Due to its superior performance, the Nvidia Jetson Nano was selected as the computational unit. Deep learning algorithms such as faster region-based convolutional neural network (R-CNN), single shot detector (SSD)- MobileNet, and single shot detector (SSD)-Inception was utilized and compared for detection and classification. From the results, the object detection algorithms using Jetson Nano achieved 80 FPS, with up to 82.98% accuracy, making it feasible for online and real-time in-vehicle monitoring with low power requirements</p>

Hierarchies in Visual Pathway: Functions and Inspired Artificial Vision.

The development of artificial intelligence has posed a challenge to machine vision based on conventional complementary metal-oxide semiconductor (CMOS) circuits owing to its high latency and inefficient power consumption originating from the data shuffling between memory and computation units. Gaining more insights into the function of every part of the visual pathway for visual perception can bring the capabilities of machine vision in terms of robustness and generality. Hardware acceleration of more energy-efficient and biorealistic artificial vision highly necessitates neuromorphic devices and circuits that are able to mimic the function of each part of the visual pathway. In this paper, wereview the structure and function of the entire class of visual neurons from the retina to the primate visual cortex within reach (Chapter 2) are reviewed. Based on the extraction of biological principles, the recent hardware-implemented visual neurons located in different parts of the visual pathway are discussed in detail in Chapters 3 and 4. Furthermore, valuable applications of inspired artificial vision in different scenarios (Chapter 5) are provided. The functional description of the visual pathway and its inspired neuromorphic devices/circuits are expected to provide valuable insights for the design of next-generation artificial visual perception systems.

Design and analysis of hybrid 10T adder for low power applications

Recently, a growing focus has been on the need for ultra-low-voltage (ULV) functioning to reduce energy usage. The full adder is a fundamental computational arithmetic unit in various image and signal processing applications. This study presents a novel architectural design for a 1-bit adder that utilizes 10 transistors. The concept relies on an energy-efficient hybrid logic multiplexer approach incorporating Gate Diffusion Input (GDI) logic. The primary goal of the suggested design for a hybrid full adder is to decrease power consumption while concurrently decreasing the required area. The suggested innovative architectural design for a full adder incorporates level restoration carry logic and ensures a complete swing output voltage. The designs have been analyzed in a standard environment using 45-nm CMOS process technology, with various supply voltages utilized. The evaluation of the adder cells focuses on speed, power consumption, power-delay-product (PDP), and space efficiency compared to the existing full adders. The recommended full adder cells were subjected to comprehensive simulation tests utilizing the Mentor Graphics environment. The results suggest that these cells surpass their counterparts, exhibiting a minimum of 80 % enhancements in their power-delay product (PDP) parameters. Also the comparisons have been made between the proposed full adders with and without level restoration techniques. The proposed level restoration technique based full adder yields better performance in terms of power and delay when compared to proposed full adder without level restoration.

Time series quantum classifiers with amplitude embedding

Quantum Machine Learning was born during the past decade as the intersection of Quantum Computing and Machine Learning. Today, advances in quantum computer hardware and the design of simulation frameworks able to run quantum algorithms in classic computers make it possible to extend classic artificial intelligence models to a quantum environment. Despite these achievements, several questions regarding the whole quantum machine learning pipeline remain unanswered, for instance the problem of classical data representation on quantum hardware, or the methodologies for designing and evaluating quantum models for common learning tasks such as classification, function approximation, clustering, etc. These problems become even more difficult to solve in the case of Time Series processing, where the context of past historical data may influence the behavior of the decision-making model. In this piece of research, we address the problem of Time Series classification using quantum models, and propose an efficient and compact representation of time series in quantum data using amplitude embedding. The proposal is capable of representing a time series of length n in log2(n)\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$log_2(n)$$\\end{document} computational units, and experiments conducted on benchmark time series classification problems show that quantum models designed for classification can also outperform the accuracy of classic methods.

Exploring machine learning to hardware implementations for large data rate x-ray instrumentation

Over the past decade, innovations in radiation and photonic detectors considerably improved their resolution, pixel density, sensitivity, and sampling rate, which all contribute to increased data generation rates. This huge data increases the amount of storage required, as well as the cabling between the source and the storage units. To overcome this problem, edge machine learning (EdgeML) proposes to move computation units near the detectors, utilizing machine learning (ML) models to emulate non-linear mathematical relationships between detector’s output data. ML algorithms can be implemented in digital circuits, such as application-specific integrated circuits and field-programmable gate arrays, which support both parallelization and pipelining. EdgeML has both the benefits of edge computing and ML models to compress data near the detectors. This paper explores the currently available tool-flows designed to translate software ML algorithms to digital circuits near the edge. The main focus is on tool-flows that provide a diverse range of supported models, optimization techniques, and compression methods. We compare their accessibility, performance, and ease of use, and compare them for two high data-rate instrumentation applications: (1) CookieBox, and (2) billion-pixel camera.