32584 research outputs found
Sort by
Assessing Perceived Safety of Non Motorized Travel with Virtual Reality
Get PDFCycling is increasingly advocated as a healthier and more sustainable mode of transportation, as reflected in both scholarly literature and policy initiatives. Nonetheless, the escalation in bicyclist crash fatalities underscores deficiencies in extant roadway designs that inadequately safeguard these vulnerable users. A persistent challenge in the examination of bicyclist safety, behavior, and comfort is the paucity of comprehensive cycling data. To enhance understanding of cyclists\u27 behavioral and physiological responses safely and efficiently, this investigation utilizes a bicycle simulator within an immersive virtual environment (IVE). Off-the-shelf sensors are employed to evaluate cyclists\u27 performance metrics (speed and lane position) and physiological responses (eye tracking and heart rate). Participants navigate a virtual environment scaled to resemble a real-world street with a shared bike lane (sharrows) to evaluate how the introduction of a curbside bike lane and a protected bike lane with flexible delineators affects perceptions of safety, as well as cyclists\u27 behavior and physiological responses. Data collected from 50 participants (across a diverse range of ages and genders) indicate that the protected bike lane design received the highest safety ratings and resulted in the lowest average cycling speed. Both the curbside bike lane and protected bike lane scenarios manifested in more focused gaze patterns than the sharrows scenario, implying that cyclists were more concentrated on the task when biking in designated lanes rather than sharing the roadway with vehicles. Furthermore, heart rate data imply that dedicated bike lanes(curbside or protected) may alleviate cyclists’ stress levels. Female participants exhibited a particular preference for the protected bike lane design
Image-to-Text Transcription: Analyzing and Describing Visual Content
No full textImage captioning, which provides a textual understanding of visual content, is the fundamental support for the advancement of Human-A.I. Interaction technology. In the hope of exploring the application of such technology, this project focuses on two specific goals. One is to directly explore the application of the image informationretrieving abilities, and the other is to dive into the specifics of the pipeline and components of image captioning models. As a result, this project presents a working app that exploits the text retrieval functionalities to enable image storage with functions like tagging and transcription. It also supports search functionality with a clear usage flow and model deployment which facilitates data security. This project also includes exploration of two trained models. One is based on Resnet50 and LSTM with adaptive attention mechanism and the other is a Qformer mid layer adapter with pretrained frozen ViT and GPT2
AI-Based Dynamic Spectrum Allocation Model for Wireless Network Management
No full textThe growth of wireless communication has introduced challenges in the dynamic and resource contrived space which is the efficient utilization of bandwidth and spectrum. This research presents a model for dynamic spectrum allocation with the help of Convolutional Neural Network (CNN) for feature extraction and the Deep Q-Network (DQN) model’s reinforcement learning architecture. The CNN captures both spatial and temporal features of the network states and gives them to the DQN for optimal allocation decision making. This CNN-DQN architecture effectively implements spectrum resource allocation in wireless networks and adapts to resource allocation changes within performance bounds. The system’s performance is evaluated in spectrum resource allocation accuracy within a loss factor, Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) showed where the allocation accuracy could still be improved. This will open the door to developing more sophisticated solutions for efficient spectrum resource management with the aid of machine learning approaches
Multilingual Sentiment Analysis Using Ensemble Learning
No full textThe widespread use of multiple social media platforms has amplified the expression of public opinions over the Internet in languages such as English, Hindi and Spanish. With the aid of technological advancements in machine learning, we can analyze opinions posted on the Internet and gauge public sentiments. There are organizations and businesses that are interested in the evaluation of these sentiments as these type of data can generally be used to obtain the opinion of a product, restaurant, a candidate, etc. In this study, we perform a comparative analysis of three popular ensemble learning methodologies (Boosting, Bagging and Stacking) based on multiple base learner models (Bidirectional Long Short-Term Memory, Support Vector Machine, Convolutional Neural Network, Gated Recurrent Unit, Recurrent Neural Network) for sentiment analysis. Publicly available multilingual datasets are used to measure the effectiveness of the models for sentiment analysis. Based on the comparative study, the results show that the stacking ensemble method produces better results than bagging and boosting to identify the correct sentiment based on multilingual data
HIERARCHICAL BLOOM FILTER TREE (HBFT): SCALABLE GEOSPATIAL METADATA INDEXING FOR BIGDATA SYSTEMS
No full textBig Data infrastructure growth has produced overwhelming metadata volumes which create extensive problems for spatial indexing and both system scalability and database queries. Traditional solutions consisting of R-trees and conventional Bloom filters manage to provide either range query support or approximate membership testing, yet they face performance issues when used at large-scale metadata management. This research proposes Hierarchical Bloom Filter Tree (HBFT) as an improved framework that integrates hierarchical spatial partitioning with partitioned, scalable, cuckoo, and striped Bloom filter variants based on existing studies in hierarchical and probabilistic indexing. The complete evaluation process shows that HBFT outperforms PostGIS (an industry-standard spatial database system). The experimental outcome indicates that HBFT delivers point, range, spatial, and aggregate queries with 5-7X performance improvement with a slight storage overhead for building the metadata index while scaling with the dataset size. The HBFT approach combines practical features of fast approximate queries with memory scalability capabilities alongside support for various query types. The combination of features in HBFT makes this system highly suitable for performing massive geospatial metadata services by supporting applications in urban planning and map-based metadata systems and high-throughput spatial analytics, which require scalable and memory-efficient querying of static data
Energy Considerations for Large Pre-trained Neural Networks
No full textIn recent years, neural network models have achieved phenomenal performance due to the increasing parameters and complexity of model architectures. However, these advancements come with high environmental costs as they require massive computational resources and consume substantial amounts of electricity, leading to high carbon emissions. Previous studies have demonstrated that substantial redundancies exist in large pre-trained models, and reducing these redundancies through compression would not compromise model performance. While these studies focused on retaining comparable model performance, the direct impact of compression on energy consumption when training models appears to have received little attention. By quantifying the energy usage associated with both uncompressed and compressed models, we investigate whether compressed models consume less electricity and validate whether compression is an effective strategy to reduce the environmental impact. Initially, we train models without compression and record electricity usage, which serves to establish a baseline for comparison. Then, as mentioned above, we consider three compression techniques: steganographic capacity reduction, pruning, and lowrank factorization. In each case, during training, we measure and record the electricity usage of each model, comparing energy consumption against model performance in terms of classification accuracy. By determining the minimum electricity required to maintain comparable model performance, this research will contribute to the development of sustainable artificial intelligence (AI) practices
AI Powered Legal Decision Support System
No full textThe large volume of legal cases presented by judicial professionals has made itchallenging to study and predict results. With advances in research methods andtechnology, predicting law cases in a more accurate manner has become an importanttrend. Prediction tools based on AI may help manage a large number of legislativetexts and documents that cannot possibly be fully read, reduce the number of casesto be seen, and give accurate outcomes of how cases may turn out. Now, whenwe look into the current AI legal prediction tools in this domain, they mostly lackefficiency and interpretability, the ability for numerous legal texts to be processed.Our AI-driven legal case prediction tool addresses these limitations, advancing thedevelopment of AI-driven legal case prediction systems that can judge the likelywinner in a lawful dispute but also explain the reason for making that decision. UsingDistilBERT-base-uncased, a large language model that we fine-tune and pre-train,we can efficiently understand legislative texts while making correct predictions aboutresults. The Supreme Court dataset from kaggle.com provides the training foundationfor our system. In order to be energy efficient, we use fine-tuning methods mainlybased on parameter fine-tuning, such as LoRA, DoRA, and QLoRA. We also exploredother LLM models like DeBERTa-v3-small, and classifiers like XGBoost. We alsoexperimented with Ex-AI, a tool to explain the interpretation of case outcomes. Ourfinal model has an accuracy of 95.6% for predicting legal cases
Large Language Model powered ETL Pipeline Development A Project Report
No full textThis projects discusses the use of Large language model to simplify and automate Extract, Transform, Loading (ETL) development. Data engineering tasks oUen need technical experDse, thereby challenging non-experts and even industry professionals for such Dme intensive operaDons. The project tackles these challenges by employing quanDzaDon of a base Lllama-2-7b-Chat model using QLora technique. The lowered precision of the base model allow further execuDon on a limited hardware resource. The project further performs Supervised fine-tuning trainer (SFT) to perform fine-tuning, specializing the model for generaDng script for ETL tasks. The fine-tuned model is evaluated through a series of micro and comprehensive end-toend ETL tasks as is compared with human-wriQen baseline scripts, demonstraDng robust performance with respect to efficiency, accuracy, and response Dme. The finetuned model achieved 85% accuracy in transformaDon tasks such as schema mapping, and data cleaning operaDons like forma`ng date, and deriving columns from raw unstructured csv files. The model is published on Hugging Face and with 81 downloads Dll date, it provides reasonable soluDon for ETL automaDon. The model also lays the foundaDon for further enhancements in the field
PHISHING DETECTION USING CONTINUAL LEARNING AND LARGE LANGUAGE MODELS
No full textAdaptive phishing detection remains crucial as the nature of cyber-attacks changes over time, which renders static models obsolete. This project extends phishing detection through the implementation of continual learning approaches, namely Elastic Weight Consolidation (EWC) and Learning Without Forgetting (LWF) with RoBERTa, a Large Language Model (LLM) and compares the results of these approaches against GPT-4o-mini, another LLM. Our approach begins with fine-tuning RoBERTa on multiple phishing datasets to establish an effective baseline. EWC is then implemented to preserve vital model parameters based on their importance measured by the Fisher Information Matrix, while LWF uses knowledge distillation to retain prior outputs when adapting to new data. On the other hand, GPT-4o-mini is evaluated through zero-shot prompting for classification and also used to generate synthetic phishing emails, addressing data scarcity and imbalance. A prototype Chrome extension is developed to demonstrate the usefulness of our adaptive system in real-time email filtering. In our experiments on four sequential phishing datasets from different time periods, RoBERTa + EWC preserved accuracies of 93%, 84%, 87% and 98% respectively, dramatically reducing catastrophic forgetting seen with static fine-tuning. RoBERTa + LWF delivered even better results, achieving accuracies of 96%, 85%, 97% and 99% on the same splits
GeneGate: Genetic Gating in a Mixture-of-Experts for Real-Time Multi-Objective Traffic Signal Control
No full textUrban traffic signal control often needs to juggle between competing goals. It needs to minimize delays, reduce emissions, prevent crashes, and prioritize emergency vehicles all while the demand is constantly fluctuating. Traditional fixed-time or statically blended policies cannot reallocate priorities quickly when conditions change. We introduce GeneGate, a mixture-of-experts framework that uses a lightweight genetic gate to fuse four specialist controllers (throughput, emissions, safety, emergency) and adjusts their weights in real time. A short offline genetic search produces a robust initial blend, and an online micro-evolution step refines it every few cycles based on live traffic feedback. GeneGate’s adaptive gating enables to detect temporal shifts in live traffic such as recognizing the intersection to be collision prone area, or emergency vehicle prone area and so on, without the need for manual retuning. In simulation on a single section testbed, GeneGate outperforms fixed time, maximum pressure, actuated baselines and a unified neural expert, delivering high reductions in collisions, travel time, and CO2 emissions, along with faster emergency response times. Furthermore, this genetic gating approach is fully general and can be applied to other dynamic environments requiring real-time, multi-objective fusion