FTT-NAS: Discovering Fault-tolerant Convolutional Neural Architecture

With the fast evolvement of embedded deep-learning computing systems, applications powered by deep learning are moving from the cloud to the edge. When deploying neural networks (NNs) onto the devices under complex environments, there are various types of possible faults: soft errors caused by cosmic radiation and radioactive impurities, voltage instability, aging, temperature variations, malicious attackers, and so on. Thus, the safety risk of deploying NNs is now drawing much attention. In this article, after the analysis of the possible faults in various types of NN accelerators, we formalize and implement various fault models from the algorithmic perspective. We propose Fault-Tolerant Neural Architecture Search (FT-NAS) to automatically discover convolutional neural network (CNN) architectures that are reliable to various faults in nowadays devices. Then, we incorporate fault-tolerant training (FTT) in the search process to achieve better results, which is referred to as FTT-NAS. Experiments on CIFAR-10 show that the discovered architectures outperform other manually designed baseline architectures significantly, with comparable or fewer floating-point operations (FLOPs) and parameters. Specifically, with the same fault settings, F-FTT-Net discovered under the feature fault model achieves an accuracy of 86.2% (VS. 68.1% achieved by MobileNet-V2), and W-FTT-Net discovered under the weight fault model achieves an accuracy of 69.6% (VS. 60.8% achieved by ResNet-18). By inspecting the discovered architectures, we find that the operation primitives, the weight quantization range, the capacity of the model, and the connection pattern have influences on the fault resilience capability of NN models.

A Consistent Geological-Seismological Model For Earthquake Occurrence in New Zealand

<p>For the development of earthquake occurrence models, historical earthquake catalogues and compilations of mapped, active faults are often used. The goal of this study is to develop new methodologies for the generation of an earthquake occurrence model for New Zealand that is consistent with both data sets. For the construction of a seismological earthquake occurrence model based on the historical earthquake record, 'adaptive kernel estimation' has been used in this study. Based on this method a technique has been introduced to filter temporal sequences (e.g. aftershocks). Finally, a test has been developed for comparing different earthquake occurrence models. It has been found that the adaptive kernel estimation with temporal sequence filtering gives the best joint fit between the earthquake catalogue and the earthquake occurrence model, and between two earthquake occurrence models obtained from data from two independent time intervals. For the development of a geological earthquake occurrence model based on fault information, earthquake source relationships (i.e. rupture length versus rupture width scaling) have been revised. It has been found that large dip-slip and strike-slip earthquakes scale differently. Using these source relationships a dynamic stochastic fault model has been introduced. Whereas earthquake hazard studies often do not allow individual fault segments to produce compound ruptures, this model allows the linking of fault segments by chance. The moment release of simulated fault ruptures has been compared with the theoretical deformation along the plate boundary. When comparing the seismological and the geological earthquake occurrence model, it has been found that a 'good' occurrence model for large dip-slip earthquakes is given by the seismological occurrence model using the Gutenberg-Richter magnitude frequency distribution. In contrast, regions dominated by long strike-slip faults produce large earthquakes but not many small earthquakes and the occurrence of earthquakes on such faults should be inferred from the dynamic fault model.</p>

A Consistent Geological-Seismological Model For Earthquake Occurrence in New Zealand

<p>For the development of earthquake occurrence models, historical earthquake catalogues and compilations of mapped, active faults are often used. The goal of this study is to develop new methodologies for the generation of an earthquake occurrence model for New Zealand that is consistent with both data sets. For the construction of a seismological earthquake occurrence model based on the historical earthquake record, 'adaptive kernel estimation' has been used in this study. Based on this method a technique has been introduced to filter temporal sequences (e.g. aftershocks). Finally, a test has been developed for comparing different earthquake occurrence models. It has been found that the adaptive kernel estimation with temporal sequence filtering gives the best joint fit between the earthquake catalogue and the earthquake occurrence model, and between two earthquake occurrence models obtained from data from two independent time intervals. For the development of a geological earthquake occurrence model based on fault information, earthquake source relationships (i.e. rupture length versus rupture width scaling) have been revised. It has been found that large dip-slip and strike-slip earthquakes scale differently. Using these source relationships a dynamic stochastic fault model has been introduced. Whereas earthquake hazard studies often do not allow individual fault segments to produce compound ruptures, this model allows the linking of fault segments by chance. The moment release of simulated fault ruptures has been compared with the theoretical deformation along the plate boundary. When comparing the seismological and the geological earthquake occurrence model, it has been found that a 'good' occurrence model for large dip-slip earthquakes is given by the seismological occurrence model using the Gutenberg-Richter magnitude frequency distribution. In contrast, regions dominated by long strike-slip faults produce large earthquakes but not many small earthquakes and the occurrence of earthquakes on such faults should be inferred from the dynamic fault model.</p>

Demystifying Unexpected Silicon Responses through User-Defined Fault Models (UDFM) and Failure Analysis

Failure Analysis (FA) plays an important role during silicon development and yield ramp up, helping identify critical test, design marginality and process issues in a timely and efficient manner. FA techniques typically rely on diagnosis callouts as a starting point for debug. Diagnostic algorithms rely on the error logs collected on production patterns which are generated to detect Stuck-at Faults (SAF) and Transition Delay Faults (TDF). Typically, SAF patterns screen out the static defects and TDF patterns test for transient fails. But often, we see cases where a SAF pattern shmoo is clean but the TDF pattern shmoo is a gross failure indicating a cell-internal static defect missed by the traditional SAF patterns. In this work, we will present our own developed User-Defined Fault Model, which targets cell-internal faults to explain unexpected silicon observations. An added advantage of the work can be seen in improving diagnosis results on the error logs collected using these targeted UDFM patterns. Since UDFM utilizes targeted fault excitation, the diagnosis algorithm results in better callouts. In this paper, we will also propose a custom diagnosis flow using our in-house UDFM to achieve better resolution. Three FA case studies will be presented to showcase the usefulness and effectivity of the proposed methods.

Grounding Fault Model of Low Voltage Direct Current Supply and Utilization System for Analyzing the System Grounding Fault Characteristics

Grounding fault analysis is of vital importance for low voltage direct current (LVDC) supply and utilization systems. However, the existing DC grounding fault model is inappropriate for LVDC supply and utilization system. In order to provide an appropriate assessment model for the DC grounding fault impact on the LVDC supply and utilization system, an LVDC supply and utilization system grounding fault model is proposed in this paper. Firstly, the model is derived by utilizing capacitor current and voltage as the system state variable, which considers the impact of the converter switch state on the topology of the fault circuit. The variation of system state parameters under various fault conditions can be easily obtained by inputting system state data in normal conditions as the initial value. Then, a model solution algorithm for the proposed model is utilized to calculated the maximum fault current, the system maximum fault current with different grounding resistance is simple to acquired based on the solution algorithm. The calculation results demonstrate that grounding resistance and structure of LVDC supply and utilization system have remarkable impacts on the transient current. The effectiveness of the proposed model is verified in PSCAD/EMTDC. The simulation results indicate that the proposed method is appropriate for the system fault analysis under various fault conditions with different grounding resistance and the proposed model can offer theoretical guidance for system fault protection.