The technology of autonomous driving vehicles has made rapid progress over the last decade, but the challenge of proving these system's safety still exists. Compared to conventional mile-based testing, scenario-based testing (SBT) is a more promising solution since scenarios covering diverse and rare driving conditions in real traffic can be simulated to assess the system's performance in safety-critical scenarios. Furthermore, understanding run-time scenarios is vital to trigger safety mechanisms designed for the Safety Of The Intended Functionality (SOTIF). However, a challenging task is to extract or generate scenarios in the design phase and recognize driving scenarios in the run-time phase due to the complexity and diversity of driving scenarios, especially the interaction with other driving agents. In this study, we proposed a complete framework for offline extraction and online identification of all kinds of interaction scenarios. A covariance-clustering based method was adopted to identify the meta-driving actions, which used the Toeplitz matrix to achieve more interpretable clustering results than distance-based methods. Experiments demonstrated the effectiveness of our method by its robust identification results for cut-in and merge-in scenarios. With a lightweight design and a theoretically valid confidence estimation method, our approach is computationally efficient for reliable online applications.

Today deep recommendation models have become increasingly large, with parameter sizes reaching hundreds of GB or even TB scale. As a result, it requires large-scale cluster computing resources to train such models. However, large-scale computing clusters tend to experience frequent failures during runtime, so fault-tolerance mechanisms such as checkpointing and restart are widely used in model training. Traditional checkpointing techniques periodically save all parameters of model, resulting in significant overhead. To address this issue, we propose an improved partial checkpointing mechanism for recommendation models named SSDC. SSDC uses an adaptive threshold strategy to reduce expensive operations when saving checkpoints, thereby having good scalability. Furthermore, SSDC saves the differential value of the model parameters, making it feasible to sparsify the otherwise dense embedding tables, thus reducing the bandwidth and time overhead to reconstruct checkpoints. Our evaluations show that compared to state-of-the-art methods, SSDC greatly reduces the time overhead of saving and reconstructing checkpoints, while achieving comparable training accuracy.

Proactive drive failure prediction can help operators handle the failing drives in advance, enhancing the storage system dependability. SSD and HDD failure prediction techniques are currently evolving towards a semi-supervised approach. In this paper, we are dedicated to enhancing the methodology for semi-supervised drive failure prediction from these aspects: design more powerful, robust yet generic models, mine drive failure modes and make the prediction model interpretable. Specifically, we propose two semi-supervised drive failure prediction models, DFP-VL and DFP-FL, from the perspective of reconstructing SMART data and learning SMART data's probability density, respectively. They capture the pattern of healthy drives, and failing drives can be detected when they deviate from the normal pattern. We mine two failure modes, slow deterioration and dramatic deterioration. Based on that, we design two failure detectors, ”ThresholdDetector” and ”HybridDetector”, to determine whether a drive deviates from the normal pattern. We evaluate the proposed methods on both SSD and HDD SMART data. DFP-VL is generic yet effective and can interpret the predicted results. DFP-FL has better performance than DFP-VL, but it cannot interpret the results. ThresholdDetector has low complexity and can detect most failing drives. HybridDetector has high complexity and can further improve the detection performance.

AdaBoost approaches have been used for multi-class imbalance classification with an imbalance ratio measured on class sizes. However, such ratio would assign each training sample of the same class with the same weight, thus failing to reflect the data distribution within a class. We propose to incorporate the density information of training samples into the class imbalance ratio so that samples of the same class could have different weights. As one could use the entire training set to calculate the imbalance and density factors, the weight of a training sample resulting from the two factors remains static throughout the training epochs. However, static weights could not reflect the up-to-date training status of base learners. To deal with this, we propose to design an adaptive weighting mechanism by making use of up-to-date training status to further alleviate the multi-class imbalance issue. Ultimately, we incorporate the class imbalance ratio, the density-based factor, and the adaptive weighting mechanism into a single variable, based on which the adaptive weights of all training samples are computed. Experimental studies are carried out to investigate the effectiveness of the proposed approach and each of the three components in dealing with multi-class imbalance classification problem.

With the ever-increasing computational demand of DNN training workloads, distributed training has been widely adopted. A combination of data, model and pipeline parallelism strategy, called hybrid parallelism distributed training, is imported to tackle the problem of deploying large-scale models. However, how to evaluate the hybrid strategy and the utilization of each device remains a challenge since existing works either profile on a real large-scale cluster with high time and money costs or only analyze a specific type of parallelism without considering the hybrid parallelism. In this work, we proposed DistSim, an event-based performance model to accurately analyze each device's computation and communication activities with low profiling costs. DistDim breaks down the model into events according to the given distributed strategy, which can be profiled on two nodes. Then DistSim leverages the hierarchy of different parallel strategies to generate the computation and communication event-flow from layer level to model level and finally the activity timeline of each device participating in training. Experiment shows that DistSim can reach <4% errors when predicting distributing training batch time and <5% errors when predicting a single device's activity time in various hybrid strategy settings. We also provide a use-case of DistSim, automatically evaluate and search the best distributed training strategy, and find a hybrid strategy with at most 7.37× throughput improvement.

Neural network robustness has been a hot research topic since the discovery of adversarial examples. With the increasing threats posed by adversarial examples in safety-critical AI applications, the research problem of how to boost the robustness of neural networks has drawn even more attention in recent years. In this paper, we argue that existing methods for neural network robustness optimization primarily focus on increasing the minimum adversarial perturbation of individual datum while neglecting the purpose of the target machine learning task, i.e., classification. Consequently, we propose a neural network retraining method which implicitly enhances a model's capability in maximizing the minimum distance from data instances of all classes to the decision boundary. As demonstrated by the experiments, our method succeeds in improving the robustness of pre-trained neural networks within only a few retraining epochs while imposing little negative effect on classification accuracy, thus being applicable for online optimization of neural networks.

As a proactive failure tolerant mechanism in large scale cloud storage systems, drive failure prediction can be used to protect data by early warning before real failures of drives, and therefore improve system dependability and cloud storage service quality. At present, solid state drives (SSDs) are generally widely used in cloud storage systems due to their high performance. SSD failures seriously affect the dependability of the system and the quality of service. Existing proactive failure tolerant mechanisms for storage systems are basically aimed at HDD failure detection and use classification technology (Supervised learning), which relies on enough failure data to establish a classification model. However, the low failure rate of SSDs leads to a serious imbalance in the ratio of positive and negative samples, which brings a big challenge for establishing a proactive failure tolerance mechanism for SSDs storage systems by using classification technology.In this paper, we propose a proactive failure tolerance mechanism for SSDs storage systems based on unsupervised technology. It only uses data of normal SSDs to train the failure prediction model, which means that our method is not limited by the imbalance in SSDs data. At the core of our method is the idea to use VAE-LSTM to learn the pattern of normal SSDs, in which case faulty SSDs can be alerted when their patterns are very different from normal ones. Our method can provide early warning of failures, thereby effectively protecting data and improving the quality of cloud storage service. We also propose a drive failure cause location mechanism, which can help operators analyze the modes of failure by providing guiding suggestions. In order to evaluate the effectiveness of our method, we use cross-validation and online testing methods on SSDs data from a technology company. The results show that the FDR and FAR of our method outperform the baselines by 17.25% and 2.39% on average.