Search results
(61 - 76 of 76)
Pages
- Title
- Exploiting contextual information for deep learning based object detection
- Creator
- Zhang, Chen
- Date
- 2020
- Description
-
Object detection has long been an important research topic in computer vision area. It forms the basis of many applications. Despite the great...
Show moreObject detection has long been an important research topic in computer vision area. It forms the basis of many applications. Despite the great progress made in recent years, object detection is still a challenging task. One of the keys to improving the performance of object detection is to utilize the contextual information from the image itself or from a video sequence. Contextual information is defined as the interrelated condition in which something exists or occurs. In object detection, such interrelated condition can be related background/surroundings, support from image segmentation task, and the existence of the object in the temporal domain for video-based object detection. In this thesis, we propose multiple methods to exploit contextual information to improve the performance of object detection from images and videos.First, we focus on exploiting spatial contextual information in still-image based object detection, where each image is treated independently. Our research focuses on extracting contextual information using different approaches, which includes recurrent convolutional layer with feature concatenation (RCL-FC), 3-D recurrent neural networks (3-D RNN), and location-aware deformable convolution. Second, we focus on exploiting pixel-level contextual information from a related computer vision task, namely image segmentation. Our research focuses on applying a weakly-supervised auxiliary multi-label segmentation network to improve the performance of object detection without increasing the inference time. Finally, we focus on video object detection, where the temporal contextual information between video frames are exploited. Our first research involves modeling short-term temporal contextual information using optical flow and modeling long-term temporal contextual information using convLSTM. Another research focuses on creating a two-path convLSTM pyramid to handle multi-scale temporal contextual information for dealing with the change in object's scale. Our last work is the event-aware convLSTM that forces convLSTM to learn about the event that causes the performance to drop in a video sequence.
Show less
- Title
- ENLARGED PERIVASCULAR SPACES IN COMMUNITY-BASED OLDER ADULTS
- Creator
- Javierre Petit, Carles
- Date
- 2020
- Description
-
Enlarged perivascular spaces (EPVS) have been associated with aging, increased stroke risk, decreased cognitive function and vascular dementia...
Show moreEnlarged perivascular spaces (EPVS) have been associated with aging, increased stroke risk, decreased cognitive function and vascular dementia. Furthermore, recent studies have investigated the links of EPVS with the glymphatic system (GS), since perivascular spaces are thought to play a major role as the main channels for clearance of interstitial solutes from the brain. However, the relationship of EPVS with age-related neuropathologies is not well understood. Therefore, more conclusive studies are needed to elucidate specific relationships between EPVS and neuropathologies. After demonstration of their neuropathologic correlates, detailed assessment of EPVS severity could provide as a potential biomarker for specific neuropathologies.In this dissertation, our focus was twofold: to develop a fully automatic EPVS segmentation model via deep learning with a set of guidelines for model optimization, and to evaluate both manual and automatic assessment of EPVS severity to investigate the neuropathologic correlates of EPVS, and their contribution to cognitive decline, by combining ex-vivo brain magnetic resonance imaging (MRI) and pathology (from autopsy) in a large community-based cohort of older adults. This project was structured as follows. First, a manual approach was used to assess neuropathologic and cognitive correlates of EPVS burden in a large dataset of community-dwelling older adults. MR images from each participant were rated using a semiquantitative 4-level rating scale, and a group of identified EPVS was histologically evaluated. Two groups of participants in descending order of average cognitive impairment were defined based and studied. Elasticnet regularized ordinal logistic regression was used to assess the neuropathologic correlates of EPVS burden in each group, and linear mixed effects models were used to investigate the associations of EPVS burden with cognitive decline. Second, a fully automatic EPVS segmentation model was implemented via deep learning (DL) using a small dataset of 10 manually segmented brain MR images. Multiple techniques were evaluated to optimize performance, mainly by implementing strategies to reduce model overfitting. The final segmentation model was evaluated in an independent test set and the performance was validated with an expert radiologist. Third, the DL segmentation model was used to segment and quantify EPVS. Quantified EPVS (qEPVS) were evaluated by combining ex-vivo MRI, pathology, and longitudinal cognitive evaluation. EPVS quantification allowed to study qEPVS both in the whole brain and regionally. Two different qEPVS metrics were studied. Elastic-net regularized linear regression was used to assess the neuropathologic correlates of qEPVS within each region of interest (ROI) under study, and linear mixed effects models were used to investigate the associations of qEPVS with cognitive decline. Finally, a preliminary study investigated the longitudinal associations of qEPVS with time. The DL segmentation model was re-trained using 4 in-vivo MR images. EPVS were segmented and quantified in a large longitudinal cohort where each participant was imaged at multiple timepoints. Factors that influenced segmentation performance across timepoints were evaluated, and linear mixed effects models controlling for these factors were used to investigate the associations of qEPVS with time.
Show less
- Title
- BIG DATA AS A SERVICE WITH PRIVACY AND SECURITY
- Creator
- Hou, Jiahui
- Date
- 2020
- Description
-
With the increase of data production sources like IoT devices (e.g., smartwatches, smartphones) and data from smart home (health sensor,...
Show moreWith the increase of data production sources like IoT devices (e.g., smartwatches, smartphones) and data from smart home (health sensor, energy sensors), truly mind-boggling amounts of data are generated daily. Building a big data as a service system, that combines big data technologies and cloud computing, will enhance the huge value of big data and tremendously boost the economic growth in various areas. Big data as a service has evolved into a booming market, but with the emergence of larger privacy and security challenges. Privacy and security concerns limit the development of big data as a service and increasingly become one of the main reasons why most data are not shared and well utilized. This dissertation aims to build a new incrementally deployable middleware for the current and future big data as a service eco-system in order to guarantee privacy and security. This middleware will retain privacy and security in the data querying and ensure privacy preservation in data analysis. In addition, emerging cloud computing contributes to providing valuable services associated with machine learning (ML) techniques. We consider privacy issues in both traditional queries and ML queries (i.e., ML classification) in this dissertation. The final goal is to design and develop a demonstrable system that can be deployed in the big data as a service system in order to guarantee the privacy of data/ service owners as well as users, enabling secure data analysis and services.Firstly, we consider a private dataset composed of a set of individuals, and the data is outsourced to a remote cloud server. We revisit the classic query auditing problem in the outsourcing scenario. Secondly, we study privacy preserving neural network classification where source data is randomly partitioned. Thirdly, we concern the privacy of confidential training dataset and models which are typically trained in a centralized cloud server but publicly accessible, \ie online ML-as-a-Service (MLaaS). Lastly, we consider the offline MLaaS systems. We design, implement, and evaluate a secure ML framework to enable MLaaS on clients' edge devices, where a ``encrypted'' ML models are stored locally.
Show less
- Title
- Systematic Serendipity: A Study in Discovering Anomalous Astrophysics
- Creator
- Giles, Daniel K
- Date
- 2020
- Description
-
In the present era of large scale surveys, big data presents new challenges to the discovery process for anomalous data. Advances in astronomy...
Show moreIn the present era of large scale surveys, big data presents new challenges to the discovery process for anomalous data. Advances in astronomy are often driven by serendipitous discoveries. Such data can be indicative of systematic errors, extreme (or rare) forms of known phenomena, or most interestingly, truly novel phenomena which exhibit as-of-yet unobserved behaviors. As survey astronomy continues to grow, the size and complexity of astronomical databases will increase, and the ability of astronomers to manually scour data and make such discoveries decreases. In this work, we introduce a machine learning-based method to identify anomalies in large datasets to facilitate such discoveries, and apply this method to long cadence light curves from NASA's Kepler Mission. Our method clusters data based on density, identifying anomalies as data that lie outside of dense regions in a derived feature space. First we present a proof-of-concept case study and we test our method on four quarters of the Kepler long cadence light curves. We use Kepler's most notorious anomaly, Boyajian's Star (KIC 8462852), as a rare `ground truth' for testing outlier identification to verify that objects of genuine scientific interest are included among the identified anomalies. Additionally, we report the full list of identified anomalies for these quarters, and present a sample subset of identified outliers that includes unusual phenomena, objects that are rare in the Kepler field, and data artifacts. By identifying <4% of each quarter as outlying data, under 6k individual targets for the dataset used, we demonstrate that this anomaly detection method can create a more targeted approach in searching for rare and novel phenomena.We further present an outlier scoring methodology to provide a framework of prioritization of the most potentially interesting anomalies. We have developed a data mining method based on k-Nearest Neighbor distance in feature space to efficiently identify the most anomalous light curves. We test variations of this method including using principal components of the feature space, removing select features, the effect of the choice of k, and scoring to subset samples. We evaluate the performance of our scoring on known object classes and find that our scoring consistently scores rare (<1000) object classes higher than common classes, meaning that rarer objects are successfully prioritized over common objects. The most common class, categorized as miscellaneous stars without any major variability, and rotational variables compose well over two-thirds of the KIC, yet are considerably underrepresented in the top outliers. We have applied scoring to all long cadence light curves of quarters 1 to 17 of Kepler's prime mission and present outlier scores for all 2.8 million light curves for the roughly 200k objects.
Show less
- Title
- Multimodal Learning and Generation Toward a Multisensory and Creative AI System
- Creator
- Zhu, Ye
- Date
- 2023
- Description
-
We are perceiving and communicating with the world in a multisensory manner, where different information sources are sophisticatedly processed...
Show moreWe are perceiving and communicating with the world in a multisensory manner, where different information sources are sophisticatedly processed and interpreted by separate parts of the human brain to constitute a complex, yet harmonious and unified intelligent system. To endow the machines with true intelligence, multimodal machine learning that incorporates data from various modalities including vision, audio, and text, has become an increasingly popular research area with emerging technical advances in recent years. Under the context of multimodal learning, the creativity to generate and synthesize novel and meaningful data is a critical criterion to assess machine intelligence.As a step towards a multisensory and creative AI system, we study the problem of multimodal generation in this thesis by exploring the field from multiple perspectives. Firstly, we analyze different data modalities in a comprehensive manner by comparing the data natures, the semantics, and their corresponding mainstream technical designs. We then propose to investigate three multimodal generation application scenarios, namely text generation from visual data, audio generation from visual data, and visual generation from textual data, with diverse approaches to give an overview of the field. For the direction of text generation from visual data, we study a novel multimodal task in which the model is expected to summarize a given video with textual descriptions, under a challenging condition where the video can only be partially seen. We propose to supplement the missing visual information via a dialogue interaction and introduce QA-Cooperative network with a dynamic dialogue history update learning mechanism to tackle the challenge. For the direction of audio generation from visual data, we present a new multimodal task that aims to generate music for a given silent dance video clip. Unlike most existing conditional music generation works that generate specific types of mono-instrumental sounds using symbolic audio representations (e.g., MIDI), and that heavily rely on pre-defined musical synthesizers, we generate dance music in complex styles (e.g., pop, breaking, etc.) by employing a Vector-Quantized (VQ) audio representation via our proposed Dance2Music-GAN (D2M-GAN) framework. For the direction of visual generation from textual data, we tackle a key desideratum in conditional synthesis, which is to achieve high correspondence between the conditioning input and generated output using the state-of-the-art generative model -- Diffusion Probabilistic Model. While most existing methods learn such relationships implicitly, by incorporating the prior into the variational lower bound in model training. In this work, we take a different route by explicitly enhancing input-output connections by maximizing their mutual information, which is achieved by our proposed Conditional Discrete Contrastive Diffusion (CDCD) framework. For each direction, we conduct extensive experiments on multiple multimodal datasets and demonstrate that all of our proposed frameworks are able to effectively and substantially improve task performance in their corresponding contexts.
Show less
- Title
- Heterogeneous Workloads Study towards Large-scale Interconnect Network Simulation
- Creator
- Wang, Xin
- Date
- 2023
- Description
-
High-bandwidth, low-latency interconnect networks play a key role in the design of modern high- performance computing (HPC) systems. The ever...
Show moreHigh-bandwidth, low-latency interconnect networks play a key role in the design of modern high- performance computing (HPC) systems. The ever-increasing need for higher bandwidth and higher message rate has driven the design of low-diameter interconnect topologies like variants of dragonfly. As these hierarchical networks become increasingly dominant, interference caused by resource sharing can lead to significant network congestion and performance variability. Meanwhile, with the rapid growth of the machine learning applications, the workloads of future HPC systems are anticipated to be a mix of scientific simulation, big data analytics, and machine learning applications. However, little work has been conducted to understand performance implications of co-running heterogeneous workloads on large-scale dragonfly systems. There is a greater need to study how different interconnect technologies affect workload performance, and how conventional scientific applications interact with emerging big data applications at the underlying interconnect level. In this work, we firstly present a comparative analysis exploring the communication interference for traditional HPC applications by analyzing the trade-off between localizing communication and balancing network traffic. We conduct trace-based simulations for applications with different communication patterns, using multiple job placement policies and routing mechanisms. Then we develop a scalable workload manager that provides an automatic framework to facilitate hybrid workload simulation. We investigate various hybrid workloads and navigate various application-system configurations for a deeper understanding of performance implications of a diverse mix of workloads on current and future supercomputers. Finally, we propose a scalable framework, Union+, that enables simulation of communication and I/O simultaneously. By combining different levels of abstraction, Union+ is able to efficiently co-model the communication and I/O traffic on HPC systems that equipped with flash-based storage. We conduct experiments with different system configurations, showing how Union+ can help system designers to assess the usefulness of future technologies in next-generation HPC machines.
Show less
- Title
- Scalable Indexing and Search in High-End Computing Systems
- Creator
- Orhean, Alexandru Iulian
- Date
- 2023
- Description
-
Rapid advances in digital sensors, networks, storage, and computation coupled with decreasing costs is leading to the creation of huge...
Show moreRapid advances in digital sensors, networks, storage, and computation coupled with decreasing costs is leading to the creation of huge collections of data. Increasing data volumes, particularly in science and engineering, has resulted in the widespread adoption of parallel and distributed file systems for storing and accessing data efficiently. However, as file system sizes and the amount of data ``owned” by users grows, it is increasingly difficult to discover and locate data amongst the petabytes of data. While much research effort has focused on methods to efficiently store and process data, there has been relatively little focus on methods to efficiently explore, index, and search data using the same high-performance storage and compute systems. Users of large file systems either invest significant resources to implement specialized data catalogs for accessing and searching data, or resort to software tools that were not designed to exploit modern hardware. While it is now trivial to quickly discover websites from the billions of websites accessible on the Internet, it remains surprisingly difficult for users to search for data on large-scale storage systems. We initially explored the prospect of using existing search engine building blocks (e.g. CLucene) to integrate search in a high-performance distributed file system (e.g. FusionFS), by proposing and building the FusionDex system, a distributed indexing and query model for unstructured data. We found indexing performance to be orders of magnitude slower than theoretical speeds we could achieve in raw storage input and output, and sought to investigate a new clean-slate design for high-performance indexing and search.We proposed the SCANNS indexing framework to address the problem of efficiently indexing data in high-end systems, characterized by many-core architectures, with multiple NUMA nodes and multiple PCIe NVMe storage devices. We designed SCANNS as a single-node framework that can be used as a building block for implementing high-performance indexed search engines, where the software architecture of the framework is scalable by design. The indexing pipeline is exposed and allows easy modification and tuning, enabling SCANNS to saturate storage, memory and compute resources on different hardware. The proposed indexing framework uses a novel tokenizer and inverted index design to achieve high performance improvement both in terms of indexing and in terms of search latency. Given the large amounts and the variety of data found in scientific large-scale file systems, it stands to reason to try to bridge the gap between various data representations and to build and provide a more uniform search space. ScienceSearch is a search infrastructure for scientific data that uses machine learning to automate the creation of metadata tags from different data sources, such as published papers, proposals, images and file system structure. ScienceSearch is a production system that is deployed on a container service platform at NERSC and provides search over data obtained from NCEM. We conducted a performance evaluation of the ScienceSearch infrastructure focusing on scalability trends in order to better understand the implications of performing search over an index built from the generated tags. Drawing from the insights gained from SCANNS and the performance evaluation of ScienceSearch, we explored the problems of efficiently building and searching persistent indexes that do not fit into main memory. The SCIPIS framework builds on top of SCANNS and further optimizes the inverted index design and indexing pipeline, by exposing new tuning parameters that allows the user to further adapt the index to the characteristics of the input data. The proposed framework allows the user to quickly build a persistent index and to efficiently run TFIDF queries over the built index. We evaluated SCIPIS over three kinds of datasets (logs, scientific data, and file system metadata) and showed that it achieves high indexing and search performance and good scalability across all datasets.
Show less
- Title
- Evaluating Speech Separation Through Pre-Trained Deep Neural Network Models
- Creator
- Prabhakar, Deeksha
- Date
- 2023
- Description
-
Speaker separation involves separating individual speakers from a mixture of voices or background noise, known as the "cocktail party problem....
Show moreSpeaker separation involves separating individual speakers from a mixture of voices or background noise, known as the "cocktail party problem." This refers to the ability to focus on a specific sound while filtering out other distractions.In this analysis, we propose the idea of obtaining features present in the original data and then evaluating the impact they have on the ability of the model to separate the mixed audio streams. The dataset is prepared such that these feature values can be used as predictor variables to various models like Logistic Regression, Decision Trees, SVM (both rbf and linear kernel), XGBoost, AdaBoost, to obtain the most contributing features that is the features that will lead to a better separation. These results shall then be analyzed to conclude the features that affect separating the audio streams the most. Initially, 400 audio streams are selected from the VoxCeleb dataset and combined to form 200 single utterances. After the mixes are obtained, the pre-trained Speechbrain model, sepformer-whamr is used. This model separates the audio mixes given as input and obtain two outputs that should be as close as possible to the original ones. A feature list from the 400 chosen audios is obtained and then the effect of certain features on the model's capability to distinguish between multiple audio sources in a mixed recording is assessed. Two analysis parameters- permutation feature importance and SHAP values are used to conclude which features have more effect on separation. Our hypothesis is that the features contributing the most to a good separation are invariant across datasets. To test this hypothesis, we obtain 1,000 audio streams from the Mozilla Common Voice Dataset and perform the same experimental methodology described above. Our results demonstrate that the features we extract from VoxCeleb dataset are indeed invariant and aid in separating the audio streams of the Mozilla Common Voice dataset.
Show less
- Title
- Heterogeneous Workloads Study towards Large-scale Interconnect Network Simulation
- Creator
- Wang, Xin
- Date
- 2023
- Description
-
High-bandwidth, low-latency interconnect networks play a key role in the design of modern high- performance computing (HPC) systems. The ever...
Show moreHigh-bandwidth, low-latency interconnect networks play a key role in the design of modern high- performance computing (HPC) systems. The ever-increasing need for higher bandwidth and higher message rate has driven the design of low-diameter interconnect topologies like variants of dragonfly. As these hierarchical networks become increasingly dominant, interference caused by resource sharing can lead to significant network congestion and performance variability. Meanwhile, with the rapid growth of the machine learning applications, the workloads of future HPC systems are anticipated to be a mix of scientific simulation, big data analytics, and machine learning applications. However, little work has been conducted to understand performance implications of co-running heterogeneous workloads on large-scale dragonfly systems. There is a greater need to study how different interconnect technologies affect workload performance, and how conventional scientific applications interact with emerging big data applications at the underlying interconnect level. In this work, we firstly present a comparative analysis exploring the communication interference for traditional HPC applications by analyzing the trade-off between localizing communication and balancing network traffic. We conduct trace-based simulations for applications with different communication patterns, using multiple job placement policies and routing mechanisms. Then we develop a scalable workload manager that provides an automatic framework to facilitate hybrid workload simulation. We investigate various hybrid workloads and navigate various application-system configurations for a deeper understanding of performance implications of a diverse mix of workloads on current and future supercomputers. Finally, we propose a scalable framework, Union+, that enables simulation of communication and I/O simultaneously. By combining different levels of abstraction, Union+ is able to efficiently co-model the communication and I/O traffic on HPC systems that equipped with flash-based storage. We conduct experiments with different system configurations, showing how Union+ can help system designers to assess the usefulness of future technologies in next-generation HPC machines.
Show less
- Title
- Utilizing Concurrent Data Accesses for Data-Driven and AI Applications
- Creator
- Lu, Xiaoyang
- Date
- 2024
- Description
-
In the evolving landscape of data-driven and AI applications, the imperative for reducing data access delay has never been more critical,...
Show moreIn the evolving landscape of data-driven and AI applications, the imperative for reducing data access delay has never been more critical, especially as these applications increasingly underpin modern daily life. Traditionally, architectural optimizations in computing systems have concentrated on data locality, utilizing temporal and spatial locality to enhance data access performance by maximizing data and data block reuse. However, as poor locality is a common characteristic of data-driven and AI applications, utilizing data access concurrency emerges as a promising avenue to optimize the performance of evolving data-driven and AI application workloads.This dissertation advocates utilizing concurrent data accesses to enhance performance in data-driven and AI applications, addressing a significant research gap in the integration of data concurrency for performance improvement. It introduces a suite of innovative case studies, including a prefetching framework that dynamically adjusts aggressiveness based on data concurrency, a cache partitioning framework that balances application demands with concurrency, a concurrency-aware cache management framework to reduce costly cache misses, a holistic cache management framework that considers both data locality and concurrency to fine-tune decisions, and an accelerator design for sparse matrix multiplication that optimizes adaptive execution flow and incorporates concurrency-aware cache optimizations.Our comprehensive evaluations demonstrate that the implemented concurrency-aware frameworks significantly enhance the performance of data-driven and AI applications by leveraging data access concurrency.Specifically, our prefetch framework boosts performance by 17.3%, our cache partitioning framework surpasses locality-based approaches by 15.5%, and our cache management framework achieves a 10.3% performance increase over prior works. Furthermore, our holistic cache management framework enhances performance further, achieving a 13.7% speedup. Additionally, our sparse matrix multiplication accelerator outperforms existing accelerators by a factor of 2.1.As optimizing data locality in data-driven and AI applications becomes increasingly challenging, this dissertation demonstrates that utilizing concurrency can still yield significant performance enhancements, offering new insights and actionable examples for the field. This dissertation not only bridges the identified research gap but also establishes a foundation for further exploration of the full potential of concurrency in data-driven and AI applications and architectures, aiming at fulfilling the evolving performance demands of modern and future computing systems.
Show less
- Title
- A Multi-level Data Integration Approach for the Convergence of HPC and Big Data Systems
- Creator
- Feng, Kun
- Date
- 2020
- Description
-
HPC is moving towards exascale (10^18 operations per second) following the trend that has continued for over half a century. Such an extremely...
Show moreHPC is moving towards exascale (10^18 operations per second) following the trend that has continued for over half a century. Such an extremely compelling computing power brings huge opportunities for scientists to explore their problems with larger sizes and finer granularity. As a result, the data volume produced and consumed by extreme-scale computing has increased dramatically. To gain useful scientific insights, scientists analyze tremendous amounts of data, which stresses the storage systems and requires efficient data access. Besides the data volume increase, the variety of I/O subsystems grows as well to meet the drastically different, often conflicting I/O requirements of numerous applications. HPC and BD, as two major camps of extreme-scale computing, have been developed separately for a long time and diverged from computing and storage paradigms. However, recent developments have proven the convergence of them leads to more efficient scientific output. Hence, unification between these ecosystems is necessary to accelerate extreme-scale computing with the collaboration of applications from both camps. Therefore, integrated I/O has become a major issue that needs to be addressed as the extreme computing community moves forward.This study explores improvement by proposing a new integrated data access system for extreme-scale computing. We enhance the BD framework to adapt to the change of integrated data access requirement by enabling direct processing of scientific data from PFS at the HPC site. Our framework can perform up to 8x faster than the state-of-the-art solutions in representative workloads. We design a new advanced I/O middleware service to utilize data aggregation resources to facilitate integrated data access in scientific workflows with both HPC and BD applications. Our middleware service can reach up to 10x speedup against the default solution and 133% better performance than existing solutions. We propose a novel storage integration solution on the storage side to unite all the storage resources, to unify the namespace across all the storage systems, and provide an ultimate integrated data access service. The integrated solution can speed up a real workflow with integrated data access requirements by up to 6.86x over existing solutions. The three-level integration at the application level, middleware level, and storage level provide us a systematic hierarchical I/O integration. Our implementation results show that the three-level optimized design and implementation is feasible and effective. It improves the state-of-the-art solutions and helps us to achieve an enhanced I/O system towards extreme-scale computing to support both HPC and BD applications.
Show less
- Title
- Resilience Enhancement of Critical Cyber-Physical Systems with Advanced Network Control
- Creator
- Liu, Xin
- Date
- 2020
- Description
-
Critical infrastructures are the systems whose failures would have a debilitating impact on national security, economics, public health or...
Show moreCritical infrastructures are the systems whose failures would have a debilitating impact on national security, economics, public health or safety, or any combination of those matters. It is important to improve those systems' resilience, which is the ability to reduce the magnitude and/or duration of disruptive events. However, today’s critical infrastructures, such as electrical power system and transportation system, are deploying advanced control applications with increasing scale and complexity, which leads to the migration of their underlying communication infrastructures from simple and proprietary networks to off-the-shelf network technologies (e.g., IP-based protocols and standards) to handle the intensive and heterogeneous traffic flows. On one hand, this migration provides an opportunity for both academic and industry communities to develop novel ideas on top of existing schemes; on the other hand, it exposes more vulnerabilities for cyber-attacks. Moreover, since the large-scale power system may choose leased networks from Internet service providers (which is a critical infrastructure itself), there exists an interdependency relationship between power and communication infrastructures, where the power transmission control requires message delivery services while the network devices rely on the power supply. These problems raise research challenges to improve the system resilience of critical cyber-physical systems.In this thesis, we focus on resilience enhancement of critical infrastructures from the communication network's aspects. The application domain includes both power and transportation systems. For power systems, we first apply advanced network control techniques (i.e., software-defined network (SDN) and fibbing control scheme) in the transmission grid communication network to improve the grid status restoration process under network failures and cyber-attacks. We develop a unified system model that contains both transmission grid monitoring system (i.e., phasor measurement unit (PMU) network) and communication network, and formalize a mixed-integer linear programming (MILP) problem to minimize the recovery time of system observability with the power and communication domain constraints. We evaluate the system performance regarding the recovery plan generation and installation using IEEE standard systems. However, the advanced network-based control scheme could also lead to problems, since it requires a power supply for the network devices. Thus, we investigate the interdependency relationship between the power grid and communication network and its impact on system resilience. We conduct a survey work that summarizes existing research based on two dimensions: objectives (i.e., failure analysis, vulnerability analysis, failure mitigation, and failure recovery) and methodologies (i.e., analytical solutions, co-simulation, and empirical studies). We also identify the limitations of existing works and propose potential research opportunities in this demanding area. Lastly, based on the review work, we conduct research that focuses on fast power distribution system restoration that involves interdependency constraints. When a natural disaster happens, both power and communication components might be damaged. Furthermore, since they are dependent on each other's service to function correctly, the failures may propagate to the hardware/software that are not affected initially. In this work, we focus on the recovery stage where the failed components in the system are already fully detected and isolated. We construct a mathematical model of the co-existing power and communication system and use optimization techniques to produce a crew dispatch plan that restores power as fast as possible by coordinating damage repairing, switch operation, and communication supply processes. We evaluate the restoration efficiency on the IEEE standard system using both analytical analysis and discrete-event simulation.For the second application domain, railway transportation system, we focus on evaluating the resilience of its communication system that exchanges control and monitoring messages with both on-board driver cabin and remote control center. We use advanced discrete-event simulation techniques to achieve a high-fidelity model of the network which makes the evaluation more concrete and realistic. For the Ethernet-based on-board train communication network (TCN), we develop a parallel simulation platform according to the IEC standard and use it to conduct a case study of a double-tagging VLAN attack on this control network. Another component of the railway communication system is the train-to-ground network that enables the communication between the driving system on the train and the control center that issues commands such as the movement authority messages. We customize the NS3 network simulator to model the LTE-based protocol with a real high-speed train trace dataset from public sources. We evaluate the resilience of the cellular network specifically on the handover process, which happens when the train travels from one base station to another. Due to the high-speed nature, the handover success rate is impacted and there are many protocol-based solutions proposed in this research area. We use the high-fidelity simulation model to evaluate some of them and compare the pros and cons.
Show less
- Title
- Efficient and Practical Cluster Scheduling for High Performance Computing
- Creator
- Li, Boyang
- Date
- 2023
- Description
-
Cluster scheduling plays a crucial role in the high-performance computing (HPC) area. It is responsible for allocating resources and...
Show moreCluster scheduling plays a crucial role in the high-performance computing (HPC) area. It is responsible for allocating resources and determining the order in which jobs are executed. Existing HPC job schedulers typically leverage simpleheuristics to schedule jobs, but such scheduling policies struggle to keep pace with modern changes and technology trends. The study of this dissertation is motivated by two new trends in HPC community: the rapid growth of heterogeneous system infrastructure and the emergence of artificial intelligence (AI) technologies. First, existing scheduling policies are solely CPU-centric. In contrast, systems become more complex and heterogeneous, and emerging workloads have diverse resource requirements, such as CPU, burst buffer, power, network bandwidth, and so on. Second, previous heuristic scheduling approaches are manually designed. Such a manual design process prevents adaptive and informative scheduling decisions. A recent trend in HPC is to intertwine AI to better leverage the investment of supercomputers. This embrace of AI provides opportunities to design more intelligent scheduling methods. In this dissertation, we propose an efficient and practical cluster scheduling framework for HPC systems. Our framework leverages AI technologies and considers system heterogeneity. The framework comprises four major components. First, shared network systems such as dragonfly-based systems are vulnerable to performance variability due to network sharing. To mitigate workload interference on these shared network systems, we explore a dedicated scheduling policy. Next, emerging workloads in HPC have diverse resource requirements instead of being CPU-centric. To cater to this, we design an intelligent scheduling agent for multi-resource scheduling in HPC leveraging the advanced multi-objective reinforcement learning (MORL) algorithm. Subsequently, we address the issues with existing state encoding approaches in RL-driven scheduling, which either lack critical scheduling information or suffer from poor scalability. To this end, we present an efficient and scalable encoding model. Lastly, the lack of interpretability of RL methods poses a significant challenge to deploying RL-driven scheduling in production systems. In response, we provide a simple, deterministic, and easily understandable model for interpreting RL-driven scheduling. The proposed models and algorithms are evaluated with real job traces from production supercomputers. Experimental results show our schemes can effectively improve job scheduling in terms of both user satisfaction and system utilization.
Show less
- Title
- Machine Learning (ML) for Extreme Weather Power Outage Forecasting in Power Distribution Networks
- Creator
- Bahrami, Anahita
- Date
- 2023
- Description
-
The Midwest region experiences a diverse range of severe weather conditions throughout the year. During the warmer months, thunderstorms,...
Show moreThe Midwest region experiences a diverse range of severe weather conditions throughout the year. During the warmer months, thunderstorms, heavy rain, lightning, tornadoes, and high winds pose a threat, while the colder season brings ice storms, snowstorms, high winds, and sleet storms, all of which can cause significant damage to the environment, properties, transportation systems, and power grids. The average climate in the Midwest is influenced by factors such as latitude, solar input, water systems' typical positions and movements, topography, the Great Lakes, and human activities. The combination of these conditions during different seasons contributes to the development of various types of storms. Therefore, it is crucial to predict the impacts of such atmospheric events on distribution and transmission lines, enabling utilities to assess and implement preventive measures and strategies to minimize the economic losses associated with these disasters. Additionally, the accurate classification of storm modes through an automated system allows operators to study trends in relation to climate change and implement necessary strategies to ensure grid reliability and resilience.In recent years, a significant number of power outages have occurred due to extreme ice formation on transmission and distribution networks, posing a threat to the power grid's resilience and reliability. To prepare power providers for snowstorms, extensive research has been conducted on snow accretion on power lines. Over the past two decades, many scientists have turned to machine learning (ML) algorithms for predicting ice accretion on overhead conductors, as ML models demonstrate superior accuracy compared to statistical forecasting models when it comes to forecasting challenging and fine-grained problems. However, most existing models primarily focus on predicting ice formation on power lines and fail to forecast the resulting damage to the distribution network. Therefore, this project proposes a model for predicting power outages caused by snow and ice storms in the distribution network. The goal is to aid in the planning process for disaster response and ensure the resilience and reliability of the power grid. The proposed outage prediction model incorporates statistical and machine learning techniques, taking into account features related to weather conditions, storm events, and information about the power network feeders.
Show less
- Title
- Utilizing Concurrent Data Accesses for Data-Driven and AI Applications
- Creator
- Lu, Xiaoyang
- Date
- 2024
- Description
-
In the evolving landscape of data-driven and AI applications, the imperative for reducing data access delay has never been more critical,...
Show moreIn the evolving landscape of data-driven and AI applications, the imperative for reducing data access delay has never been more critical, especially as these applications increasingly underpin modern daily life. Traditionally, architectural optimizations in computing systems have concentrated on data locality, utilizing temporal and spatial locality to enhance data access performance by maximizing data and data block reuse. However, as poor locality is a common characteristic of data-driven and AI applications, utilizing data access concurrency emerges as a promising avenue to optimize the performance of evolving data-driven and AI application workloads.This dissertation advocates utilizing concurrent data accesses to enhance performance in data-driven and AI applications, addressing a significant research gap in the integration of data concurrency for performance improvement. It introduces a suite of innovative case studies, including a prefetching framework that dynamically adjusts aggressiveness based on data concurrency, a cache partitioning framework that balances application demands with concurrency, a concurrency-aware cache management framework to reduce costly cache misses, a holistic cache management framework that considers both data locality and concurrency to fine-tune decisions, and an accelerator design for sparse matrix multiplication that optimizes adaptive execution flow and incorporates concurrency-aware cache optimizations.Our comprehensive evaluations demonstrate that the implemented concurrency-aware frameworks significantly enhance the performance of data-driven and AI applications by leveraging data access concurrency.Specifically, our prefetch framework boosts performance by 17.3%, our cache partitioning framework surpasses locality-based approaches by 15.5%, and our cache management framework achieves a 10.3% performance increase over prior works. Furthermore, our holistic cache management framework enhances performance further, achieving a 13.7% speedup. Additionally, our sparse matrix multiplication accelerator outperforms existing accelerators by a factor of 2.1.As optimizing data locality in data-driven and AI applications becomes increasingly challenging, this dissertation demonstrates that utilizing concurrency can still yield significant performance enhancements, offering new insights and actionable examples for the field. This dissertation not only bridges the identified research gap but also establishes a foundation for further exploration of the full potential of concurrency in data-driven and AI applications and architectures, aiming at fulfilling the evolving performance demands of modern and future computing systems.
Show less
- Title
- Multimodal Learning and Generation Toward a Multisensory and Creative AI System
- Creator
- Zhu, Ye
- Date
- 2023
- Description
-
We are perceiving and communicating with the world in a multisensory manner, where different information sources are sophisticatedly processed...
Show moreWe are perceiving and communicating with the world in a multisensory manner, where different information sources are sophisticatedly processed and interpreted by separate parts of the human brain to constitute a complex, yet harmonious and unified intelligent system. To endow the machines with true intelligence, multimodal machine learning that incorporates data from various modalities including vision, audio, and text, has become an increasingly popular research area with emerging technical advances in recent years. Under the context of multimodal learning, the creativity to generate and synthesize novel and meaningful data is a critical criterion to assess machine intelligence.As a step towards a multisensory and creative AI system, we study the problem of multimodal generation in this thesis by exploring the field from multiple perspectives. Firstly, we analyze different data modalities in a comprehensive manner by comparing the data natures, the semantics, and their corresponding mainstream technical designs. We then propose to investigate three multimodal generation application scenarios, namely text generation from visual data, audio generation from visual data, and visual generation from textual data, with diverse approaches to give an overview of the field. For the direction of text generation from visual data, we study a novel multimodal task in which the model is expected to summarize a given video with textual descriptions, under a challenging condition where the video can only be partially seen. We propose to supplement the missing visual information via a dialogue interaction and introduce QA-Cooperative network with a dynamic dialogue history update learning mechanism to tackle the challenge. For the direction of audio generation from visual data, we present a new multimodal task that aims to generate music for a given silent dance video clip. Unlike most existing conditional music generation works that generate specific types of mono-instrumental sounds using symbolic audio representations (e.g., MIDI), and that heavily rely on pre-defined musical synthesizers, we generate dance music in complex styles (e.g., pop, breaking, etc.) by employing a Vector-Quantized (VQ) audio representation via our proposed Dance2Music-GAN (D2M-GAN) framework. For the direction of visual generation from textual data, we tackle a key desideratum in conditional synthesis, which is to achieve high correspondence between the conditioning input and generated output using the state-of-the-art generative model -- Diffusion Probabilistic Model. While most existing methods learn such relationships implicitly, by incorporating the prior into the variational lower bound in model training. In this work, we take a different route by explicitly enhancing input-output connections by maximizing their mutual information, which is achieved by our proposed Conditional Discrete Contrastive Diffusion (CDCD) framework. For each direction, we conduct extensive experiments on multiple multimodal datasets and demonstrate that all of our proposed frameworks are able to effectively and substantially improve task performance in their corresponding contexts.
Show less