YorkSpace has migrated to a new version of its software. Access our Help Resources to learn how to use the refreshed site. Contact diginit@yorku.ca if you have any questions about the migration.

Computer Science

Permanent URI for this collection

https://hdl.handle.net/10315/38508

Browse

Now showing 1 - 20 of 70

Open Access
Enhancing code review for improved code quality with language model-driven approaches
(2024-03-16) Rahman, Shadikur; Prince, Enamul Hoque
Code review is essential for maintaining software development standards, yet achieving effective reviews and issue resolution remains challenging. This thesis introduces RefineCode, an application tool to find actionable code reviews and provide similar code reviews as references within an organization, aiding developers in resolving issues effectively. To this end, we collected 9,500 code reviews from five private projects in an industrial setting and empirically evaluated various classification methods for identifying actionable code reviews. RefineCode automatically recommends relevant solutions from Stack Overflow based on textual similarity and entity linking between code reviews and Stack Overflow issues. Additionally, it integrates a chatbot feature, leveraging large language models to propose potential solutions for actionable code reviews. These features empower developers to make informed decisions, enhancing code quality by guiding issue resolution without reinforcing misunderstandings.
Open Access
Towards Efficient and Robust Caching: Investigating Alternative Machine Learning Approaches for Edge Caching
(2024-03-16) Torabi, Hoda ; Litoiu, Marin; Khazaei, Hamzeh
This study introduces HR-Cache, a caching framework designed to enhance the efficiency of edge caching. The increasing complexity and variability of traffic classes at edge environments pose significant challenges for traditional caching methods, which often rely on simplistic metrics. HR-Cache addresses these challenges by implementing a learning-based strategy grounded in Hazard Rate ordering, a concept originally used to establish cache performance upper bounds. By employing a lightweight supervised machine learning model, HR-Cache learns from HR-based caching decisions and predicts the "cache-friendliness" of incoming requests, identifying "cache-averse" objects as priority candidates for eviction. Our experiment results demonstrate HR-Cache's superior performance. It consistently achieves 2.2–14.6% greater WAN traffic savings compared to the LRU strategy and outperforms both heuristic and state-of-the-art learning-based algorithms, while adding minimal prediction overhead. Though designed with the considerations of edge caching limitations, HR-Cache can be adapted with minimal changes for broader applicability in various caching contexts.
Open Access
Trajectory Prediction Learning using Deep Generative Models
(2024-03-16) Li, Jing; Papagelis, Manos
Trajectory prediction involves estimating an object's future path using its current state and historical data, with applications in autonomous vehicles, robotics, and human motion analysis. Deep learning methods trained on historical data have been applied to this task, but they struggle with complex spatial dependencies due to the intricate nature of trajectory data and dynamic environments. We introduce TrajLearn, a novel trajectory prediction model using generative models and higher-order mobility flow representations (hexagons). TrajLearn, given a trajectory's recent history and current state, predicts its next k steps. It employs a variant of beam search for exploring multiple paths, ensuring spatial continuity. Our experiments demonstrate that TrajLearn surpasses current leading methods and other baselines by about 60% on various real-world datasets. We also explore different prediction horizons (k values), perform resolution sensitivity analysis, and conduct an ablation study to evaluate the contributions of different model components.
Open Access
Multi-Versioning and Microservices: A Strategy for Developing Reliable Software Systems
(2024-03-16) Akhtarian, Nazanin; Khazaei, Hamzeh
In the dynamic realm of software engineering, adaptability is key to sustaining system performance and reliability. Software iterations often bring about challenges such as unexpected bugs and performance issues, necessitating a nuanced approach to maintain system integrity. In this work, we propose employing software multi-versioning to enhance system reliability. We embark on an in-depth exploration of the reliability of microservices within chaotic environments. Using Chaos Mesh, we simulate a series of disruptions in a microservices-based application, i.e., the Online Boutique. Through real experimentation, we systematically introduce various chaos disruptions, such as Pod failures, response delay, and memory stress, to investigate their impact on the system's reliability. We define a reliability metric that quantifies the robustness and efficiency of each software version under adverse conditions. Leveraging this metric, we introduce a dynamic controller that adjusts the population of each version, ensuring optimal resource distribution, reliability and system performance. Additionally, our research evaluates how the system adapts to varying workloads. We investigate how well the system can adjust its scalability—specifically, the number of replicas—in response to changes in \acrshort{cpu} usage as the user load fluctuates. Our findings demonstrates the system's capability to scale dynamically based on workload demands, ensuring robustness and efficiency. In conclusion, our study provides a detailed framework for employing software multi-versioning as a means to enhance system reliability. By devising a reliability metric and implementing a dynamic scaling system that responds to both reliability assessments and workload variations, we offer a comprehensive strategy to fortify systems against the unpredictable nature of software evolution, ensuring they remain resilient and make efficient use of resources.
Open Access
Key-Frame Based Motion Representations for Pose Sequences
(2024-03-16) Thasarathan, Harrish Patrick; Derpanis, Konstantinos
Modelling human motion is critical for computer vision tasks that aim to perceive human behaviour. Extending current learning-based approaches to successfully model long-term motions remains a challenge. Recent works rely on autoregressive methods, in which motions are modelled sequentially. These methods tend to accumulate errors, and when applied to typical motion modelling tasks, are limited up to only four seconds. We present a non-autoregressive framework to represent motion sequences as a set of learned key-frames without explicit supervision. We explore continuous and discrete generative frameworks for this task and design a key-framing transformer architecture to distill a motion sequence into key-frames and their relative placements in time. We validate our learned key-frame placement approach with a naive uniform placement strategy and further compare key-frame distillation using our transformer architecture with an alternative common sequence modelling approach. We demonstrate the effectiveness of our method by reconstructing motions up to 12 seconds.
Open Access
Data Acquisition for Domain Adaptation of Closed-Box Models
(2024-03-16) Liu, Yiwei; Yu, Xiaohui
Machine learning (ML) marketplace provides customers with various ML solutions to accelerate their business. Models in the ML market are often available as closed boxes, but they may suffer from distribution shifts in new domains. Prior techniques cannot address this problem, because they are either impractical to use or against the property of closed-box models. Instead, we propose to acquire extra data to construct a "padding" model to help the original closed box with its classification weaknesses in the target domain. Our solution consists of a "weakness detector" to discover the deficiency of the original closed-box model and the Augmented Ensemble approach to combine the source and the padding model for better performance in the target domain and further diversifying the ML marketplace. Extensive experiments on several popular benchmark datasets confirm the superiority and robustness of our proposed framework over baseline approaches.
Open Access
Examining Autoexposure for Challenging Scenes
(2024-03-16) Yang, Beixuan; Brown, Michael S.
Autoexposure (AE) is a critical step cameras apply to ensure properly exposed images. While current AE algorithms are effective in well-lit environments with unchanging illumination, these algorithms still struggle in environments with bright light sources or scenes with abrupt changes in lighting. A significant hurdle in developing new AE algorithms for challenging environments, especially those with time-varying lighting, is the lack of platforms to evaluate AE algorithms and suitable image datasets. To address this issue, we have designed a software platform allowing AE algorithms to be used in a plug-and-play manner with the dataset. In addition, we have captured a new 4D exposure dataset that provides a complete solution space (i.e., all possible exposures) over a temporal sequence with moving objects, bright lights, and varying lighting. Our dataset and associate platform enable repeatable evaluation of different AE algorithms and provide a much-needed starting point to develop better AE methods.
Open Access
Advancing Blind Face Restoration: Robustness and Identity Preservation with Integrated GAN and Codebook Prior Architectures
(2024-03-16) Tayarani Bathaie, Seyed Nima ; An, Aijun
Blind Face Restoration (BFR) is a challenging task in computer vision, which aims to reconstruct High-Quality (HQ) facial images from Low-Quality (LQ) inputs. BFR presents as a challenging ill-posed problem, necessitating auxiliary information to constrain the solution space. While geometric and generative facial priors provide some support in BFR, their effectiveness wanes under intense degradation. Discrete codebook priors, though promising, grapple with the difficulty of associating intensely degraded images with their corresponding codes. To effectively address these limitations, this research introduces a two-stage restoration approach, termed Identity-embedded GAN and Codebook Priors (IGCP), which synergistically combines the strengths of both generative and codebook priors. In the first stage, our approach employs a Generative Prior Restorer (GPR) network for initial image restoration. Distinct from existing methods that apply identity-based losses to the final restored image, our work innovates by embedding identity information directly into the style vectors of the StyleGAN2 network during the generation process. This is achieved through the introduction of an \emph{identity-in-style} loss, ensuring superior fidelity and identity preservation even in severely degraded images Proceeding to the second stage, the approach utilizes a two-component framework known as the Codebook Prior Restorer (CPR) network. This framework comprises a Vector Quantized AutoEncoder (VQAE) for artifact mitigation and to add a final touch of quality, complemented by introducing a Feature Transfer Module (FTM) that is demonstrated to be necessary to ensure fidelity and identity preservation. Extensive experimental evaluations were conducted across five datasets, including our newly introduced CelebA-IntenseTest dataset. The results from these experiments demonstrate the remarkable efficacy of the IGCP approach. Notably, IGCP has shown exceptional performance in handling various degradation levels, setting new benchmarks in the domain of BFR.
Open Access
Active Visual Search: Investigating human strategies and how they compare to computational models
(2024-03-16) Wu, Tiffany; Tsotsos, John K.
Real world visual search by fully active observers has not been sufficiently investigated. Whilst the visual search paradigm has been widely used, most studies use a 2D, passive observation task, where immobile subjects search through stimuli on a screen. Computational models have similarly been compared to human performance only to the degree of 2D image search. I conduct an active search experiment in a 3D environment, measuring eye and head movements of untethered subjects during search. Results show patterns forming strategies for search, such as repeated search paths within and across subjects. Learning trends were found, but only in target present trials. Foraging models encapsulate subject location-leaving actions, whilst robotics models captured viewpoint selection behaviours. Eye movement models were less applicable to 3D search. The richness of data collected from this experiment opens many avenues of exploration, and the possibility of modelling active visual search in a more human-informed manner.
Open Access
Precision Recall Cover: A Method to Assess Generative Models
(2023-12-08) Cheema, Fasil Tariq; Urner, Ruth
Generative modelling has seen enormous practical advances over the past few years from LLMs like ChatGPT to image generation. However, evaluating the quality of a generative system is often still based on subjective human inspection. To overcome this, very recently, the research community has turned to exploring formal evaluation metrics and methods. In this work, we propose a novel evaluation method based on a two-way nearest neighbor test. We define a new measure of mutual coverage for two probability distributions. From this, we derive an empirical analogue and show analytically that it exhibits favorable theoretical properties while it is also straightforward to compute. We show that, while algorithmically simple, our derived method is also statistically sound. We complement our analysis with a systematic experimental evaluation and comparison to other recently proposed measures. Using a wide array of experiments, we demonstrate our algorithm’s strengths over other existing methods and confirm our results from the theoretical analysis.
Open Access
Investigating Calibrated Classification Scores through the Lens of Interpretability
(2023-12-08) Torabian, Alireza; Urner, Ruth
Calibration is a frequently invoked concept when useful label probability estimates are required on top of classification accuracy. A calibrated model is a scoring function whose scores correctly reflect underlying label probabilities. Calibration in itself however does not imply classification accuracy, nor human interpretable estimates, nor is it straightforward to verify calibration from finite data. There is a plethora of evaluation metrics (and loss functions) that each assesses a specific aspect of a calibration model. In this work, we initiate an axiomatic study of the notion of calibration and evaluation measures for calibration. We catalogue desirable properties of calibration models as well as evaluation metrics and analyze their feasibility and correspondences. We complement this analysis with an empirical evaluation, comparing two metrics and comparing common calibration methods to employing a simple, interpretable decision tree.
Open Access
Leveraging Deep Learning for Trajectory Similarity Learning and Trajectory Pathlet Dictionary Construction
(2023-12-08) Alix, Gian Carlo Idris; Papangelis, Emmanouil
The rapid development of geospatial technologies and location-based devices have motivated the research community of trajectory data mining, due to numerous applications including route planning and navigation services. Of interest are similarity search tasks that several works addressed through representation learning. Our method ST2Box offers refined representations by first representing trajectories as sets of roads, then adapting set-to-box architectures for learning accurate, versatile, and generalizable set representations of trajectories for preserving similarity. Experimentally, ST2Box outperforms baselines by up to ~38%. Another related problem involves constructing small sets of building blocks that can represent wide-ranging trajectories (pathlet dictionaries). However, currently-existing methods in constructing PDs are memory-intensive. Thus, we propose PathletRL for generating dictionaries that offer significant memory-savings. It initializes unit-length pathlets and iteratively merges them while maximizing utility -- that is approximated using deep reinforcement learning-based method. Empirically, PathletRL can reduce its dictionary's size by up to 65.8% against state-of-the-art methods.
Open Access
Chart Question Answering with an Universal Vision-Language Pretraining Approach
(2023-12-08) Parsa Kavehzadeh; Enamul Hoque Prince
Charts are widely used for data analysis, providing visual representations and insights into complex data. To facilitate chart-based data analysis using natural language, several downstream tasks have been introduced recently including chart question answering. However, existing methods for these tasks often rely on pretraining on language or vision-language tasks, neglecting the explicit modeling of chart structures. To address this, we first build a large corpus of charts covering diverse topics and visual styles. We then present UniChart, a pretrained model for chart comprehension and reasoning. We propose several chart-specific pretraining tasks that include: (i) low-level tasks to extract the visual elements (e.g., bars, lines) and data from charts, and (ii) high-level tasks to acquire chart understanding and reasoning skills. Our experiments demonstrate that pretraining UniChart on a large corpus with chart-specific objectives, followed by fine-tuning, yields state-of-the-art performance on four downstream tasks. Moreover, our model exhibits superior generalizability to unseen chart corpus, surpassing previous approaches that lack chart-specific objectives and utilize limited chart resources.
Open Access
Evaluating Temporal Queries over Videos
(2023-12-08) Chen, Yueting; Yu, Xiaohui
Videos have been an important part of people's daily lives and are continuously growing in terms of volume, size, and variety of content. Recent advances in Computer Vision (CV) algorithms have improved accuracy and efficiency, making video annotations possible with high accuracy. In this work, we follow a general framework to first obtain annotations utilizing state-of-the-art CV algorithms, and then consider three research problems on evaluating temporal queries with such annotations. Specifically, we first investigate the temporal queries that consider only co-occurrence relationships between objects on video feeds, where we take the first step and define such queries in a way that they incorporate certain physical aspects of video capture such as object occlusion. We propose two techniques, Marked Frame Set (MFS) and Sparse State Graph (SSG), to organize all detected objects in the intermediate data generation layer, which effectively, given the queries, minimizes the number of objects and frames that have to be considered during query evaluation. Then, we consider the query with a ranking mechanism that aims to retrieve clips from large video repositories in which objects co-occur in a query-specified fashion. We propose a two-phased approach, where we build indexes during the Ingestion Phase, and then answer queries during the Query Phase using the Partition-Based Query Processing (PBQP) algorithm, which efficiently produces the desired (query-specified) number of results with the highest scores. Finally, we further consider both spatial and temporal information with graph representations and define the problem of Spatial and Temporal Constrained Ranked Retrieval (STAR Retrieval) over videos. Based on the graph representation, we propose a two-phase approach, consisting of the ingestion phase, where we construct and materialize the Graph Index (GI), and the query phase, where we compute the top-ranked windows (video clips) according to the window matching score efficiently. We propose two algorithms to perform Spatial Matching (SMA) and Temporal Matching (TM) separately with an early-stopping mechanism. We present the details of the above three research problems and our proposed methods. Via experiments conducted on various datasets, we show the effectiveness of our proposed methods.
Open Access
Fine Granularity is Critical for Intelligent Neural Network Pruning
(2023-12-08) Heyman, Andrew Baldwin; Zylberberg, Joel
Neural network pruning is a popular approach to reducing the computational costs of training and/or deploying a network, and aims to do so while minimizing accuracy loss. Pruning methods that remove individual weights (fine granularity) yield better ratios of accuracy to parameter count, while methods that preserve some or all of a network’s structure (coarser granularity, e.g. pruning channels from a CNN) take better advantage of hardware and software optimized for dense matrix computations. We compare intelligent iterative pruning using several different criteria sampled from the literature against random pruning at initialization across multiple granularities on two different image classification architectures and tasks. We find that the advantage of intelligent pruning (with any criterion) over random pruning decreases dramatically as granularity becomes coarser. Our results suggest that, compared to coarse pruning, fine pruning combined with efficient implementation of the resulting networks is a more promising direction for improving accuracy-to-cost ratios.
Open Access
A 360-degree Omnidirectional Photometer Using a Ricoh Theta Z1
(2023-12-08) MacPherson, Ian Michael; Brown, Michael S.
Spot photometers measure the luminance emitted or reflected from a small surface area in a physical environment. Because the measurement is limited to a "spot," capturing dense luminance readings for an entire environment is impractical. This thesis demonstrates the potential of using an off-the-shelf commercial camera to operate as a 360-degree luminance meter. The method uses the Ricoh Theta Z1 camera, which provides a full 360-degree omnidirectional field of view and an API to access the camera's minimally processed RAW images. Working from the RAW images, this thesis describes a calibration method to map the RAW images under different exposures and ISO settings to luminance values. By combining the calibrated sensor with multi-exposure high-dynamic-range imaging, a cost-effective mechanism for capturing dense luminance maps of environments is provided. The results show that the Ricoh Theta calibrated as a luminance meter performs well when validated against a significantly more expensive spot photometer.
Open Access
Query-Aware Data Systems Tuning via Machine Learning
(2023-12-08) Henderson, Connor Dustin; Szlichta, Jarek
Modern data systems have hundreds of system configuration parameters which heavily influence the performance of business queries. Manual configuration by experts is painstaking and time consuming. We propose a query-informed tuning system called BLUTune which uses deep reinforcement learning based on advantage actor-critic neural networks to tune configurations within defined resource constraints. We translate high-dimensional query execution plans into a low-dimensional embedding space and illustrate the usefulness of query embeddings for the downstream task of data systems tuning. We train our model based on the estimated cost of queries then fine-tune it using query execution times. We present an experimental study over various synthetic and real-world workloads. One model uses TPC-DS queries such that there are tables from the schema that are not seen during training time. The second is trained under resource constraints to show how the model performs when we limit the memory the system has access to.
Open Access
Automation in Open Source Software: A GitHub Marketplace Analysis
(2023-12-08) Saroar, Sk Golam; Nayebi, Maleknaz
This thesis comprises two papers that examine automation tools in the Open Source Software (OSS) ecosystem on GitHub, focusing on GitHub Actions as well as the GitHub Marketplace, which is a platform for sharing these Actions for collaboration and reuse. Our research aims to understand and explore the state of automation in OSS, as existing studies have mainly focused on statistical analysis of a sample of GitHub repositories, neither considering developers’ perspectives nor leveraging the GitHub Marketplace. The first paper conducted a survey analysis to investigate the motivations, decision criteria, and challenges associated with creating, publishing, and using Actions. The second paper explores the GitHub Market- place and presents a mapping study by analyzing 7,878 Actions and 515 research papers mapped into 32 different categories. We found a substantial industry-academia gap, with researchers focusing on experimentation and practitioners relying more on exploration tools. The limited number of OSS automation tools published in academia contrasted with the convenient access practitioners had to the marketplace offerings. This thesis contributes to the understanding of automation in the OSS ecosystem, highlights the industry-academia gap, offers insights for researchers to build on existing work, and aids practitioners in navigating technology and finding synergies.
Open Access
Fault Analysis and Control of DFIGs for Grid Code Compliance and Protection of Power System
(2023-12-08) Mohammadpour, Hassan; Hooshyar, Ali
Inverter-based resources (IBRs) are growing at exponential rates in today's power systems. Therefore, a sizable portion of the measurements of relays is expected to come from IBRs. However, the fault current characteristics of IBRs put the operation of the relays in jeopardy as they are different than that of synchronous generators' (SGs) based on which the relays' operating principles are developed. Therefore, different countries have progressively revised their grid codes (GCs) to reduce the likelihood of protection malfunctions and ensure stable and continuous operation of power systems. Similar to emerging regional GCs, the recently approved IEEE 2800 Standard mandates that IBRs generate negative-sequence current during low-voltage ride-through (LVRT) conditions. The 2800 Standard requires that the IBRs' negative-sequence current lead the negative-sequence voltage by 90-100 degrees to emulate SGs and reduce the likelihood of protection malfunction. However, the limitations of existing doubly-fed induction generators (DFIGs) led the Standard to exempt the DFIGs from this requirement and allow a wider range for their negative-sequence current angle. Meanwhile, the 2800 Standard also acknowledged that this exemption had unidentified and potentially negative impacts on protective relays. This dissertation, for the first time, (i) sheds light on several so-far-unknown DFIG characteristics that impact the angle of the negative-sequence current during LVRT, (ii) reveals the impacts of the above DFIG exemption on industrial relays, and (iii) develops a solution to prevent the need for this exemption in the future revisions of the IEEE 2800 Standard. This dissertation also investigates the challenges brought about by the DFIGs during the crowbar connection and rectification mode of operation, i.e., interrupted control of the DFIG's converters, now affecting the performance of distance relays that are installed at a DFIG-based wind farm substation. The focus is on the relays implemented using the apparent impedance approach and the commercially developed reactance method. It is revealed that the phase elements of a distance relay that uses these methods are prone to under-/over-reach in the systems with DFIGs. The exclusive fault behavior of DFIGs along with different units of a distance relay is scrutinized to identify the root causes. To address the relay problems, a communication-assisted method with minimal bandwidth requirement is developed, which provides non-delayed fast tripping over the entire length of the line.
Open Access
Cyber-Physical Attacks Detection and Resilience Methods in Smart Grids
(2023-12-08) Sawas, Abdullah; Farag, Hany E. Z.
Backed by the deployment of increasingly reliable Information and Communication Technologies (ICT) infrastructure, modern power systems heavily depend on computerized circuits to function within an interconnected environment. In particular, Smart Grids (SGs) core domain relies on ICTs networks and components to communicate control signals and data measurements to improve the efficiency of power generation and distribution while maintaining safe and reliable operations. ICTs have also extended the SG domain of interaction to include other utilities, such as the natural gas grid to efficiently utilize multiple energy forms and resources. In the consumer domain, a growing number of appliances and autonomous smart loads equipped with Internet of Things (IoT) technology are being deployed into SG, the results in large portions of electric demand being remotely controlled. Despite their advantages, ICTs are vulnerable to cyber–attacks that can deteriorate SGs' operational safety and integrity. Thus, new approaches to enhance the resiliency of SGs against cyber-physical attacks are needed. To that extent, this thesis develops new resiliency investigation approaches under the three aforementioned domains. First, in the SG domain, an efficient False Data Injection Attack (FDIA) approach is developed imitating an intelligent adversary behavior searching for an optimal attack vector against State Estimation (SE) modules. Simulation results show that using this approach, an adversary can identify attack vectors with minimal size and superior flexibility to manipulate, in real-time, power flow measurements of the system lines as perceived by the SE without the need to acquire additional measurements. Hence, attacks constructed using this approach require less computational time and resources compared to the existing methods making it beneficial for the analysis of cyber–security vulnerabilities and the design of resilient SE modules. Second, under the Integrated Energy System (IES) domain, an operational framework model is developed to be used as a testbed for performing and analyzing the impact of cyber–attacks. The framework models steady–state power and gas flow operations, and presents a new financial interdependency operation scheduling model. The framework is validated on standard power distribution and transmission systems with variable generation and demand scenarios and high renewable penetration levels. Using this framework, an attack resiliency method is developed based on signal processing and machine-learning tools. The method is able to detect 98.6% and 94.5% of the external signals and internal control commands respectively. Third, the vulnerability of Power Distribution Systems (PDSs) to compromised collections of IoT-enabled appliances is investigated, and a stealthy attack strategy is presented. Accordingly, a new index is developed, referred to as Feeder Loading Abnormal Power Spectrum (FLAPS), and used in a novel real-time detection and prediction approach to counter stealthy attacks and estimate the attack onset time. Results demonstrate that the method is able to detect and alert for stealthy attacks in a timely manner, thereby enabling the system to operate reliably and securely. By identifying new attacks, and proposing detection methods and countermeasures, this thesis contributes to the collective efforts to address the risks associated with cyber–attacks against the SGs components. Specifically, the quantitative results show that deploying the proposed methods will enhance the resiliency of SE and IESs, and protect the PDSs against threats of large-scale deployment of IoT-enabled appliances.