System-level Prediction Research Articles

Adopting Two Supervisors for Efficient Use of Large-Scale Remote Deep Neural Networks

Recent decades have seen the rise of large-scale Deep Neural Networks (DNNs) to achieve human-competitive performance in a variety of artificial intelligence tasks. Often consisting of hundreds of millions, if not hundreds of billion parameters, these DNNs are too large to be deployed to, or efficiently run on resource-constrained devices such as mobile phones or IoT microcontrollers. Systems relying on large-scale DNNs thus have to call the corresponding model over the network, leading to substantial costs for hosting and running the large-scale remote model, costs which are often charged on a per-use basis. In this paper, we propose BiSupervised , a novel architecture, where, before relying on a large remote DNN, a system attempts to make a prediction on a small-scale local model. A DNN supervisor monitors said prediction process and identifies easy inputs for which the local prediction can be trusted. For these inputs, the remote model does not have to be invoked, thus saving costs, while only marginally impacting the overall system accuracy. Our architecture furthermore foresees a second supervisor to monitor the remote predictions and identify inputs for which not even these can be trusted, allowing to raise an exception or run a fallback strategy instead. We evaluate the cost savings, and the ability to detect incorrectly predicted inputs on four diverse case studies: IMDB movie review sentiment classification, Github issue triaging, Imagenet image classification, and SQuADv2 free-text question answering. In all four case studies, we find that BiSupervised allows to reduce cost by at least 30% while maintaining a similar system-level prediction performance. In two case studies (IMDB and SQuADv2), we find that BiSupervised even achieves a higher system-level accuracy, at reduced cost, compared to a remote-only model. Furthermore, measurements taken on our setup indicate a large potential of BiSupervised to reduce average prediction latency.

A robust model selection framework for fault detection and system health monitoring with limited failure examples: Heterogeneous data fusion and formal sensitivity bounds

Fault detection models play a fundamental role in monitoring the health state of engineering systems subject to degradation processes. Data-driven fault detection models, albeit very effective when trained on large databases of failures, fail to perform well under a lack of failure examples. Because reliable engineering systems seldom fail, data shortage is often inevitable. To overcomes failure data scarcity, this work proposes a new model selection framework for the robust selection of fault detection and system health monitoring models. We combine heterogeneous sources of information (time-to-failure and sensors data), a model of the system structure, and mathematical bounds on the sensitivity and specificity of components fault classifiers. We use Support Vector Machines (SVMs) to detect components faults and Scenario theory to derive formal sensitivity and specificity bounds. The component predictions are combined within a system structure-function for system health states estimation. A novel model selection strategy optimizes the hyper-parameters of the SVM ensemble by minimizing system-level prediction errors and generalization error bounds of the individual fault classifiers. One of the main advantages of the proposed framework is a set of formal epistemic bounds on fault detection and false alarm probabilities quantifying the lack of data uncertainty affecting the fault detection rate. We test the method on three representative case studies: (1) On randomized fault detection experiments with synthetic data, (2) on ten SVM models for predictive maintenance of industrial health care imaging systems, and (3) on a real-world PHM challenge problem. The results prove the efficacy of the proposed approach and its usefulness for solving fault detection.

Structural Landmarking and Interaction Modelling: A “SLIM” Network for Graph Classification

Graph neural networks are a promising architecture for learning and inference with graph-structured data. Yet, how to generate informative, fixed dimensional features for graphs with varying size and topology can still be challenging. Typically, this is achieved through graph-pooling, which summarizes a graph by compressing all its nodes into a single vector. Is such a “collapsing-style” graph-pooling the only choice for graph classification? From complex system’s point of view, properties of a complex system arise largely from the interaction among its components. Therefore, we speculate that preserving the interacting relation between parts, instead of pooling them together, could benefit system level prediction. To verify this, we propose SLIM, a graph neural network model for Structural Landmarking and Interaction Modelling. The main idea is to compute a set of end-to-end optimizable sub-structure landmarks, so that any input graph can be projected onto these (spatially) local structural representatives for a faithful, global characterization. By doing so, explicit interaction between component parts of a graph can be leveraged directly in generating discriminative graph representation. Encouraging results are observed on benchmark datasets for graph classification, demonstrating the value of interaction modelling in the design of graph neural networks.

A Cross-Layer Framework for Temporal Power and Supply Noise Prediction

In modern microprocessor and SoC designs, supply noise margin has been significantly reduced due to the continuously decreasing supply voltage level. On the other hand, with increasing current density, chips may see larger supply noise variations on various spots and from time to time. As a result, chip robustness and reliability are inevitably deteriorated with more frequent supply noise emergencies. It is therefore crucial to have an efficient supply noise prediction method to enhance design robustness. The state-of-art solutions either try to build a spatial noise estimation framework at the layout-level using the limited distributed physical noise sensors or attempt to develop emergency predictors at the architecture-level thus ignore back-end power delivery details. In this paper, we propose a cross-layer framework for temporal supply noise prediction. Our method not only accounts for the temporal characteristics of workload execution at micro-architecture-level but also incorporates the power delivery model at the circuit-level into such system-level prediction. In order to enable the capability of on-the-fly noise prediction, we first bridge the gap between system-level workload and micro-architectural-level power by employing an ordinary least square-based power estimation model and an adaptive auto-regressive integrated moving average model (ARIMA)-based power prediction model. Then a layout-level supply noise model is developed to explore the correlations between micro-architectural-level power and layout-level supply noise. Compared with existing methods, the proposed ARIMA-based power model improves the prediction performance by up to 37.5%/63.0% in X86/ARM. Moreover, compared with SPICE simulation, our framework is able to estimate present supply noise with an average error of 0.005% and predict future supply noise with an average error of 1.58%/1.17% for X86/ARM architecture.

An approach for composing predictive models from disparate knowledge sources in smart manufacturing environments

This paper describes an approach that can compose predictive models from disparate knowledge sources in smart manufacturing environments. The capability to compose disparate models of individual manufacturing components with disparate knowledge sources is necessary in manufacturing industry, because this capability enables us to understand, monitor, analyze, optimize, and control the performance of the system made up of those components. It is based on the assumption that the component models and component sources used in any particular composition can be represented using the same collection of system ‘viewpoints’. With this assumption, creating this integrated collection is much easier than it would be. This composition capability provides the foundation for the ability to predict the performance of the system from the performances of its components—called compositionality. Compositionality is the key to solve decision-making/optimization problems related to that system-level prediction. For those problems, compositionality can be achieved using a three-tiered, abstraction architecture. The feasibility of this approach is demonstrated in an example in which a multi-criteria decision making method is used to determine the optimal process parameters in an additive manufacturing process.

The use of linked data to develop and validate measures of variation in mortality and readmissions for public reporting purposes in New South Wales, Australia

ABSTRACTObjectivesTo investigate variation across 78 New South Wales public hospitals, in mortality in the 30 days following admission and in returns to acute care (readmissions) in the 30 days following discharge for acute myocardial infarction, ischaemic stroke, heart failure, pneumonia, hip fracture surgery.
 ApproachLinked data were used to (1) construct an analytic unit – an index period of care that comprised concatenated acute, contiguous hospitalisations with the principal diagnosis of interest; (2) to capture outcomes both within the index hospital and following discharge, wherever they occurred; (3) to enhance risk adjustment with one year look back for relevant comorbidities; (4) to assess fair attribution of outcomes. A risk-standardised mortality ratio (RSMR) and a risk standardised readmission ratio (RSRR) were calculated as the ratio of the observed to the expected number events at a given hospital, by developing and validating condition specific system-level prediction models. Funnel plots identified outliers. For the RSRR, the competing risk of death was considered.
 ResultsFor both outcome indicators, sensitivity was enhanced by the use of linked data (33%-100% more deaths; 23%–32% more returns to acute care or readmissions). For mortality, RSMRs that only capture deaths in hospital, as opposed to deaths within 30 days of admission, were shown to be biased and change the outlier status of about 20% of hospitals. Including socioeconomic status in risk adjustment models altered the outlier status of about 10% of hospitals on the cusp of statistical significance but did not significantly alter the RSMRs. For returns to acute care, sensitivity analyses that included socioeconomic status in the models found there was no significant improvement in discriminatory power. For example, in the case of ischaemic stroke, the c-statistic for the model without inclusion of SES was 0.593 (0.578-0.610); inclusion of SES resulted in a c-statistic of 0.600 (0.583-0.616). There were some changes in hospital-level results but there was no clear evidence of a systematic effect on results.
 ConclusionThe risk-standardised ratio method, based on linked data, compares a hospital’s results given its case mix with an average New South Wales hospital with the same case mix. Ratio-based indicators have been reported publicly and have proven to be a valuable screening tool to identify hospitals where further investigation may be required locally.

Prediction of air-oil interface location in tied-shaft type fluid dynamic bearings

Tied-shaft type Fluid dynamic bearings (FDB) are widely used for small, high precision motor applications such as hard disk drives. This bearing contains two or more air-oil interfaces that cause the meniscus location to change as the bearing rotates from the stationary condition. To prevent oil shortage or leakage at the seal of the FDB, designing an internal geometry of the FDB that can maintain the airoil interface location at a suitable range is critical. This study proposes a system-level prediction model for the air-oil interface location in the taper seal element of the FDB in operating conditions. The FDB elements, grooves and a circulation hole are represented as pumps and flow resistance. With these representations, the FDB can be analyzed with a lumped parameter model. In addition, a hydrostatic equilibrium condition that incorporates capillary force at the meniscus in the taper seal element is used. The location of the meniscus is determined by the pressure difference between the upper and lower seal that is generated by the hydrodynamic effect inside the FDB in operating conditions. The characteristics of the FDB elements are obtained from CFD analysis and are used to solve five equations that predict the system-level performance characteristics: Upper/lower meniscus height, flow rate, pressure and flying height. With the proposed prediction model, the effects of the upper/lower chamfer location and the inclination angle of the circulation hole are investigated. These results confirm that an oil shortage occurs if the upper chamfer radius is below 1.5 mm, and the upper/lower chamfer radii have a bigger effect on the meniscus location than the inclination angle of the circulation hole.

Role of calibration, validation, and relevance in multi-level uncertainty integration

Calibration of model parameters is an essential step in predicting the response of a complicated system, but the lack of data at the system level makes it impossible to conduct this quantification directly. In such a situation, system model parameters are estimated using tests at lower levels of complexity which share the same model parameters with the system. For such a multi-level problem, this paper proposes a methodology to quantify the uncertainty in the system level prediction by integrating calibration, validation and sensitivity analysis at different levels. The proposed approach considers the validity of the models used for parameter estimation at lower levels, as well as the relevance at the lower level to the prediction at the system level. The model validity is evaluated using a model reliability metric, and models with multivariate output are considered. The relevance is quantified by comparing Sobol indices at the lower level and system level, thus measuring the extent to which a lower level test represents the characteristics of the system so that the calibration results can be reliably used in the system level. Finally the results of calibration, validation and relevance analysis are integrated in a roll-up method to predict the system output.

Integration of model verification, validation, and calibration for uncertainty quantification in engineering systems

This paper proposes a Bayesian methodology to integrate model verification, validation, and calibration activities for the purpose of overall uncertainty quantification in different types of engineering systems. The methodology is first developed for single-level models, and then extended to systems that are studied using multi-level models that interact with each other. Two types of interactions amongst multi-level models are considered: (1) Type-I, where the output of a lower-level model (component and/or subsystem) becomes an input to a higher level system model, and (2) Type-II, where parameters of the system model are inferred using lower-level models and tests (that describe simplified components and/or isolated physics). The various models, their inputs, parameters, and outputs, experimental data, and various sources of model error are connected through a Bayesian network. The results of calibration, verification, and validation with respect to each individual model are integrated using the principles of conditional probability and total probability, and propagated through the Bayesian network in order to quantify the overall system-level prediction uncertainty. The proposed methodology is illustrated with numerical examples that deal with heat conduction and structural dynamics.

UDSM Trends Comparison: From Technology Roadmap to UltraSparc Niagara2

The increased leakage, yield inefficiency, process, power supply, and temperature variations have significant aftereffects on the performance of complex VLSI architectures especially if mapped on ultra deep sub micrometer (UDSM) technologies. In this paper we assess the technology trend based on three industrial technologies (90, 65, and 45 nm) using a state of the art processor as benchmark: The UltraSparc Niagara 2 from SUN Microsystem. We analyze frequency, dynamic, and static power and area after synthesis varying power supply voltage and temperature. We then compare these exhaustive analyses of system level performance as a function of technology to ITRS device level estimations. The results suggest that this prediction can be of help when addressing both the technological scaling and the variability scenario of the selected technology. We believe that correctly predicting specific values on performance variations when realistic conditions and technologies are changed could provide a valuable information for the architect. Our analysis advises the designer on the effective applicability of the ITRS trends to system performance, but also pinpoints that a reliable system level prediction should better take into account the design complexity.

The current-voltage relation for electropores with conductivity gradients.

In electroporation, an electric field transiently permeabilizes the cell membrane to gain access to the cytoplasm, and to deliver active agents such as DNA, proteins, and drug molecules. Past work suggests that the permeabilization is caused by the formation of aqueous, conducting pores on the lipid membrane, which are also known as electropores. The current-voltage relation across the membrane-bound pores is critical for understanding and predicting electroporation. In this work, we solve the Nernst-Planck equations in a geometry encompassing an isolated electropore to investigate this relation. In particular, we study cases where the intra- and extracellular electrical conductivities differ. We first derive an analytical solution, which is subsequently validated with a direct numerical simulation using a finite volume method. The main result of the current work is a formula for the effective pore resistance as a function of the pore radius, the membrane thickness, and the intra- and extracellular conductivities. This formula can be incorporated into whole-cell or planar-membrane electroporation models for system-level prediction and understanding.

On the system level prediction of joint time frequency spreading systems with carrier phase noise

Phase noise is a topic of theoretical and practical interest in electronic circuits. Although progress has been made in the characterization of its description, there are still considerable gaps in its effects especially on multi-carrier spreading systems. In this paper, we investigate the impact of a local oscillator phase noise on the multi-carrier 2 dimensional (2D) spreading systems based on a combination of orthogonal frequency division multiplexing (OFDM) and code division multiple access (CDMA) and known as OFDM-CDMA. The contribution of this paper is multifold. First, we use some properties of random matrix and free probability theory to give a simplified expression of signal to interference and noise ratio (SINR) obtained after equalization and despreading. This expression is independent of the actual value of the spreading codes and depends mainly on the complex amplitudes of estimated channel coefficients. Secondly, we use this expression to derive new weighting functions which are very interesting for the radio frequency (RF) engineers when they design the frequency synthesizer. Therefore, based on these asymptotic results, we adapt a new method to predict the bit error rate (BER) at the output of the channel decoder by using an effective SINR value. We show by simulations the validity of our models and that at a given BER, the required signal to noise ratio (SNR) may easily increase due to the carrier phase noise.

An Effective Technique for System-Level Prediction of the Radiated Emissions of Unknown Sources Inside Low-${Q}$ Cavities Using Unit-Level Measurements

<para xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> A method for predicting electromagnetic emissions from unknown electronic devices and cables inside low- <formula formulatype="inline"><tex Notation="TeX">$\bm{Q}$</tex></formula> cavities is presented. This method is based on measurements at the unit level, equivalent source decomposition, and numerical electromagnetic simulation of the integrated system. The approach is also valid for the analysis of cables and conductors carrying electrical currents from the interior to the exterior of the measurement surface (Huygen's surface) and is useful for the prediction of the electromagnetic compatibility characteristics of electronic devices located inside satellites. </para>

Interconnect-based system-level energy and power prediction to guide architecture exploration

We present a novel technique to predict energy and power consumption in an electronic system, given its behavioral specification and library components. The early prediction gives circuit designers the freedom to make numerous high-level choices (such as die size, package type, and latency of the pipeline) with confidence that the final implementation will meet power and energy as well as cost and performance constraints. Our unique statistical estimation technique associates low-level, technology dependent physical and electrical parameters, with expected circuit resources and interconnect. Further correlations with switching activity yield accurate results consistent with implementations. All feasible designs are investigated using this technique and the designer may tradeoff between small size, high speed, low energy, and low power. The results for designs of two popular signal processing applications, predicted prior to synthesis, are within 10% accuracy of power estimates performed on synthesized layouts.

2513 LMS Virtual. Lab : 次世代 CAE プラットフォーム

In the past, virtual prototyping technologies simply did not deliver the right answers in time to impact the ongoing design process. Very often, CAE has been used for design verification and has run almost outside the process - not as a proactive design optimization tool. Raw processing speed has not been the real bottleneck. The fundamental problem lies with the lack of accurate system-level prediction technologies and with the sheer number of different software tools that are in use at many companies today. A lot of time is lost in non value-added engineering effort - dealing with many barely compatible software tools for just one attribute at a time, transferring data from one environment to another, rebuilding models, and remeshing.

Prioritizing system-reliability prediction improvements

This method prioritizes system-reliability prediction activities once a preliminary reliability-prediction has been made. System-reliability predictions often use data and models from a variety of sources, each with differing degrees of estimation uncertainty. Since time and budgetary constraints limit the extent of analyzes and testing needed to estimate component reliability, it is necessary to allocate limited resources intelligently. A reliability-prediction prioritization index (RPPI) is defined to provide a relative ranking of components based on their potential for improving the accuracy of a system-level reliability prediction by decreasing the variance of the system-reliability estimate. If a component has a high RPPI, then additional testing or analysis should be considered to decrease the variance of the component reliability estimate. RPPI is based on a decomposition of the variance of the system-reliability or on a mean-time-to-failure estimate. Using these indexes, the effect of individual components within the system can be compared, ranked, and assigned to priority groups. The ranking is based on whether a decrease of the component-reliability estimate variance meaningfully decreases the system-reliability estimate variance. The procedure is demonstrated with two examples.