Physical Sciences and Mathematics Commons^™

Toward Mapping Multiphase Multicomponent Mixtures With Neural Networks, Kristen L. Hallas, Melissa De Jesus, Christine J. Wu, Jianzhi Li, Jason Bernstein, Philip C. Myint

School of Mathematical & Statistical Sciences Faculty Publications

Equation of state (EOS) tables are commonly used in hydrodynamic simulations of high-pressure, high-temperature phenomena in fields like planetary science, astrophysics, and high-energy-density science. However, generating and storing EOS tables for multiphase, multicomponent mixtures over a wide range of pressures and temperatures is computationally infeasible due to their memory-intensive nature. To address this issue, we have developed a neural network-based machine learning model to predict new EOS tables for binary mixtures. In particular, a deep feedforward neural network trained on a set of ten EOS tables at particular mixture compositions is able to predict nine new (hold-out) EOS tables at …

Port And Vessel Communication Traffic Intrusion Detection: A Variational Autoencoder‑Enhanced Multilayer Perceptron Approach, Chien-Lin Chiang, Hsien-Cheng Chou, Ming-Yuan Peng, Yi-Yuan Chiang, Yu-Shun Liu

Journal of Marine Science and Technology–Taiwan

Port and vessel networks increasingly operate on IP/Ethernet backbones with high‑noise, high‑dimensional traffic. We present a lightweight hybrid intrusion‑detection model that couples a variational autoencoder (VAE) with a multilayer perceptron (MLP) and augments training with a boundary‑oriented latent‑space mixup strategy. The VAE models the distribution of normal traffic and identifies anomalies through reconstruction errors. Subsequently, it generates robust latent vectors, enabling the MLP to perform highly accurate supervised classification. On the UNSW‑NB15 dataset, the proposed pipeline attains ≥97% accuracy and an outstanding recall of 99.56% in binary intrusion detection, and visualization of the latent space (PCA) together with reconstruction‑error analyses …

A Deep Learning Based Three-Stage Fusion Neural Network With Composite Adaptive Momentum Optimizer For Bangla Cyberbully Identification, Mahin Montasir Afif, Abdullah Al Noman, Kazi Abdullah Jarif, Md Emamul Arefin Islam, Sathya Narayana Sharma K, Mahfujur Rahman, Dipta Gomes, Kazi Tanvir, Md Obaidur Rahaman, Syed Mohammed Shamsul Islam

Research outputs 2022 to 2026

Cyberbullying in Bangla social media is an increasingly critical issue, posing severe risks to mental health and online safety. Despite the growing need for automated detection systems, challenges such as limited linguistic resources and the nuanced characteristics of the Bangla language hinder model performance. In this study, we propose TriB-FNN, a novel and explainable three-stage fusion neural network, designed to effectively capture cyberbully in Bangla social media texts. The model integrates Convolutional Neural Networks (CNNs), Bidirectional Long Short-Term Memory (Bi-LSTM), and Gated Recurrent Units (GRUs) in a hierarchical pipeline. To optimize training, we employ a Composite Adaptive Momentum (CAM) Optimizer, …

Fig-Gan: Fundus Image Generation Via Deep Learning Based Generative Adversarial Network For Amd Disease Diagnosis, Kailasa Thrishul, Ahed Abugabah, Amina Salhi, Manel Ayadi, Mohamed M. Sithik, D. Jayaprakash, A. Ahilan

All Works

Globally, age-related macular degeneration (AMD) remains a main cause of irreversible vision loss. Recently, deep learning models have primarily focused on classifying fundus images for early detection of AMD progression. However, existing models rarely address the generation of future progression-aware fundus images, particularly when complete real longitudinal follow-up scans are unavailable. This limitation makes it difficult to track retinal changes over time and highlights the need for generative models capable of producing realistic drusen-level structural variations. To address these issues, a novel deep learning-based FIG-GAN model is to generate synthetic future fundus images from baseline inputs. Multi-Attention U-Net (MAU-Net) is …

A Deep Learning Ensemble Framework For Multi-Subtype Renal Tumor Classification Using Contrast-Enhanced Ct, Hisham Abdeltawab, Ahmed Alksas, Mohamed Ghazal, Ashraf Khalil, Norah Saleh Alghamdi, Rasha T. Abouelkheir, Ahmed Elmahdy, Mohamed Abou El-Ghar, Sohail Contractor, Ayman El–Baz

All Works

Renal cell carcinoma (RCC) is considered the most aggressive and common form of renal cancer. Therefore, early detection is crucial to ensure appropriate and effective treatment planning. In our study, we propose a novel computer-aided diagnostic (CAD) approach which incorporates a deep learning ensemble to differentiate between five renal tumor subtypes, utilising the modality of contrast-enhanced computed tomography (CE-CT). The addressed renal lesions are malignant tumors (chromophobe RCC (chRCC), papillary RCC (pRCC), and clear cell RCC (ccRCC)) and benign tumors (renal oncocytoma (RO) and angiomyolipoma (AML)). Our study includes 280 patients who underwent renal biopsy, 112 patients were diagnosed with …

Keyframe Selection From Motion Capture Data With Dual-Agent Reinforcement Learning, Kun Hu, Wang, Clinton Mo, Mingyang Ma, Shaohui Mei, Zebin Chen, Zhiyong Wang

Research outputs 2022 to 2026

Animation production workflows centered around motion capture techniques require animators to edit motions based on a set of keyframes. However, most existing keyframe selection methods are optimization-based, which suffer from the issues of flexibility and efficiency. In this paper, a novel deep reinforcement learning method with dual agents are proposed for unsupervised keyframe selection. First, an S-Agent and an R-Agent evaluate the actions of selection and refinement, respectively. A deep spatio-temporal network, namely graph keyframe evaluation network (GKEN), is proposed for the agents. Then, an animation specified reward is devised based on reconstruction, which fulfills three important properties of the …

Optimized Fully Sharded Data Parallel For Efficient Distributed Training Of Bilstm With Multi-Head Attention For Network Intrusion Detection Systems, Noor Al-Huda H. Oleiwi, Ameer K. Hadi

Journal of Intelligent Informatics, Networking, and Cybersecurity

The expanding volume of network traffic, the growing size of cybersecurity datasets, and the increasing complexity of deep learning models present major computational challenges to training deep learning-based Network Intrusion Detection Systems (NIDSs). Fully Sharded Data Parallel (FSDP) enables distributed training by sharding model parameters across multiple GPUs; however, the synchronization and communication overhead can limit its training efficiency. To overcome these limitations, this paper proposes an Optimized Fully Sharded Data Parallel (Optimized FSDP) framework that improves distributed training efficiency by dynamic gradient accumulation and synchronization scheduling. The proposed framework is integrated with a Bidirectional Long Short-Term Memory (BiLSTM) network …

Long-Tail Object Detection Method In Transportation Scene Images Based On Improved Rt-Detr, Mingjun Deng, Peihua Nie, Junhua Guo, Ningbo Gao

Journal of System Simulation

Abstract: Detecting rare objects in transportation environments is crucial for the safety and reliability of autonomous driving systems. To address the challenges of limited detection capability on rare objects that appear only a few times during training, we propose an improved long-tailed object detection method for traffic scene images named TF-DETR (tail free DETR) based on RT-DETR model. First, the lightweight CSP-EDA (CSP-efficient depthwise aggregation) was constructed to reconstruct the backbone network. Its core unit, the efficient depthwise aggregation, integrated the pooled-transpose attention and convolutional gated linear unit to enhance feature representations for tail-class samples. Second, the spatially-adaptive dynamic cognitive …

Hybrid Attention For Low-Overlap Point Cloud Registration, Huaijun Wang, Zonghua Yu, Chenxin Xu, Junhuai Li, Chenzhi Li, Junjie Yin

Journal of System Simulation

Abstract: To address the significant registration errors caused by large initial pose deviations in low overlap ratio point cloud, this paper presents a two-stage optimization method that integrates selfattention and cross-attention mechanisms. Orientation feature representation is enhanced using equivariant self-attention, semantic alignment is achieved through invariant cross-attention, and rotational equivariance is modeled with an octahedral group for coarse registration. Various attention variants are employed to aggregate features along the anchor dimension, and a local-to-global module is utilized for fine registration. On the 3DLoMatch dataset, the proposed method achieves a registration recall of 77.0% and an inlier ratio of 44.3%. Even …

Jpnet: A Multi-Layered Fusion Deep Learning Architecture For The Detection And Classification Of Pests In Jute Crops, Mejbah Ahammad, Md. Ashraful Babu, Muhammad Sajjad Hossain, Md. Fayz-Al-Asad, Nadim Ahmed, Md. Khaled Hossain, Md. Mortuza Ahmmed, M. Mostafizur Rahman, Mufti Mahmud

Mathematical Modelling and Numerical Simulation with Applications

Detecting and classifying insect pests is a critical challenge in agricultural pest management, as infestations can reduce crop yield and quality. This study introduces JPNet, a convolutional neural network (CNN) architecture that uses multi-layer feature fusion to detect and classify insect pests affecting jute crops. The architecture integrates complementary feature representations extracted at different network depths, preserving fine-grained visual characteristics alongside high-level semantic information. JPNet is evaluated on the JutePest dataset, which comprises approximately 6,460 RGB images spanning 17 pest classes. Preprocessing and data augmentation—including resizing, normalization, rotation, shifting, zooming, and flipping—improve the consistency and diversity of the training data. …

A Fault Identification Method For Deep Coalbed Methane Reservoirs Based On An Improved Transunet, Gao Congshun, Wu Haibo, Liu Lei, Wang Zijian, Huang Xuanyu, Jia Lilan, Xu Aiguo

Coal Geology & Exploration

Objective Small-scale faults in deep coalbed methane (CBM) reservoirs significantly affect CBM enrichment and the deployment of development wells. However, deep seismic data tend to suffer from low signal-to-noise ratios (SNRs) and poor continuity of reflection events, while conventional methods are likely to miss small-scale faults. To address these limitations, this study proposes a fault identification method based on an improved TransUNet. Methods The proposed method is developed based on the R50-ViT-B_16 TransUNet architecture. A coordinate attention (CA) module is introduced before the skip connections enter the decoder to enhance the feature representation of the locations and directional continuity of …

A Deep Learning-Based Approach For Mapping Shrubs In Arctic Tundra From Very High-Resolution Imagery, Darko Radakovic, Mark Chopping, Bruce Cook, Aparna S. Varde, Stefan A. Robila

Department of Earth and Environmental Studies Faculty Scholarship and Creative Works

Arctic shrub expansion is accelerating due to climate warming, but accurately mapping it at fine spatial scales remains challenging due to spectral overlap between shrubs and surrounding vegetation, heterogeneous tundra landscapes, and limited availability of very high-resolution imagery. This study presents a novel mapping and validation approach that integrates very high-resolution (< 1 m) QuickBird (QB) and WorldView (WV) satellite imagery across 12 North Slope sites, using deep learning architectures to map changes in shrub cover from 2002 to 2022. Deep learning models, including Convolutional Neural Network (CNN), ResNet, VGG, U-Net, and Vision Transformer, were trained on 4,100 3-channel pansharpened multispectral images (200×200 m). Two traditional machine learning models (Random Forest and Support Vector Machine) served as a control group. Leave-one-site-out cross-validation showed that all models performed well (AUC-ROC of 0.78–0.86, F1 at validation of 0.33-0.46) and had a moderate to strong agreement with field reference data (R2 = 0.76–0.87). The pretrained VGG and ResNet models achieved the strongest field-reference alignments for shrub cover prediction, while the difference with traditional models was small. Temporal trend analyses for July and August demonstrate a cover change of 0.06–0.45% yr-1 with unweighted mean area slopes and 0.01–0.92% yr-1 inverse-variance weighting. Growth rates were found to depend on the analytical design. Furthermore, the limiting factor in mapping shrub cover was not model complexity but the quality of the training data labels (annotation method explained ∼26% of the error variance). Per-scene cover error was seen across architectures (ρ = 0.67). Comparison with NASA Goddard's LiDAR, Hyperspectral and Thermal (G-LiHT)-derived Canopy Height Models revealed a weak-to-moderate relationship (R2 = 0.04–0.77), suggesting horizontal cover and vertical canopy structure capture complementary aspects of shrub abundance. These results demonstrate the potential of deep learning to map decadal shrub expansion at sub-meter resolution.

A Calibrated And Conformal Deep Learning Framework For Trustworthy Antinuclear Antibody Pattern Recognition With Selective Referral To Experts, Hussein Ali Hussein Al Naffakh, Ahmed Dheyaa Radhi, Raghdah Maytham Hameed, Muntaha Abdullah Reishaan, Fouad A. Majeed, Rozaida Ghazali

Karbala International Journal of Modern Science

Reading antinuclear antibody patterns on human epithelial cells by indirect immunofluorescence is the reference screen for systemic autoimmune rheumatic diseases, but it is slow, subjective, and variable between observers. Deep learning reaches high accuracy on this task, yet most systems return a single prediction without stating how reliable it is, which is unsafe in a diagnostic workflow. This paper presents an intelligent decision support framework built around a single calibrated uncertainty signal. That signal is the control variable for four reliability modules: confidence calibration, conformal prediction, error detection, and selective referral. A feature space out of distribution detector serves as …

Dynamic Focal Loss Adjustment For Railway Defect Detection, Mehmet Koç, Ridvan Özdemi̇r, Ömer Gerek

Turkish Journal of Electrical Engineering and Computer Sciences

Railway infrastructure is critical to the safe and efficient operation of transportation systems, and the early detection of defects is essential for preventing catastrophic failures. Automated defect detection methods are therefore crucial for maintaining continuous safety while reducing maintenance costs. Although Focal Loss is widely used in object detection under class-imbalanced conditions, its fixed α parameter may limit its effectiveness in detecting rare defects. In this study, we propose an adaptive α-tuned Focal Loss approach that dynamically adjusts class weights based on average precision (AP) values. By iteratively optimizing α without relying on gradient-based optimization, the proposed method improves the …

Pahdf: A Privacy-Aware Hybrid Detection Framework With Class-Aware Weighted Stacking Ensemble (Cawse) For Fake Instagram Account Detection, Sura Jasim Mohammed, Safa Saad Abbas, Suhad Hatem Jihad

Journal of Intelligent Informatics, Networking, and Cybersecurity

The rapid growth of social media platforms has intensified concerns regarding online privacy, data security, and fraudulent activities that driven by fake accounts. This paper proposes a Privacy-Aware Hybrid Detection Framework (PAHDF) to detect Instagram fake account that integrates privacy preservation with high-performance machine learning. Unlike existing approaches that treat privacy and detection as separated objectives, therefore, the proposed framework jointly addresses both objectives by relying exclusively on publicly available, low-sensitivity profile metadata. PAHDF combines a deep learning model for latent feature representation with a Random Forest classifier for behavioural pattern learning through a Class-Aware Weighted Stacking Ensemble (CAWSE), where …

Geographical Pattern Analysis Of Gis Images With Deep Learning And Voronoi Network, Nidaa Kareem, Tawfiq A. Al-Assadi

Journal of Intelligent Informatics, Networking, and Cybersecurity

Localization is not enough for the analysis of spatial patterns; a principled geometric and statistical framework is required. This paper proposes an integrated spatial intelligence system combining deep learning, computational geometry, and spatial statistics, which is a unified and interpretable system. It is based on segmentation localization that accurately localizes the centroid of each object without the disadvantages of the bounding box. These centroids form a natural Voronoi tessellation of regions of spatial influence intrinsic to the data instead of imposing any artificial restrictions. A geometry-based density formulation is used to improve representation, which includes Voronoi cell areas and neighborhood …

The Application Of Machine Learning And Deep Learning On Demand Forecasting Across Time-Critical Industries: A Systematic Review, Asmaa Seyam, Sujith Samuel Mathew, May El Barachi, Cheng Zhang, Jun Shen

All Works

The applications of machine learning and deep learning in demand forecasting have attracted increasing attention, as they offer remarkable predictive capabilities that help automate forecasting processes and achieve higher accuracy. While numerous review studies have examined solutions within specific industries, there is a lack of comprehensive literature review investigating these solutions across different sectors. Therefore, this study overviews machine learning and deep learning applications in demand forecasting across time-critical industries, including power, tourism, water, transportation, and food. A two-tier classification framework is proposed to categorize demand forecasting studies by both application industry and methodological architecture. In addition, the most popular …

Light-Wdef: Lightweight Deepfake Evaluation Framework For Efficient And Reliable Visual Detection, Muhammad Yousaf, Sarwar Shah Khan, Babar Shah, Muhammad Shehzad Khan, Muhammad Awais Bacha, Mian Saeed Akbar, Ijaz Ali, Fernando Moreira

All Works

Deepfake technology synthesizes and manipulates digital content to produce artificial videos in which an individual’s identity, facial appearance, or actions are convincingly altered. These forgeries can be highly realistic, posing serious risks to trust in digital media and undermining personal reputation, dignity, and the integrity of public discourse. This paper presents the Lightweight Deepfake Evaluation Framework (Light-wDEF)– a resource-efficient, reproducible framework tailored for large-scale evaluation of visual deepfake detectors on modest hardware. Light-wDEF combines deterministic frame sampling, batched MTCNN face extraction, streaming append-only storage with lazy offset indexing, online Albumentations augmentations, stratified splits with class-aware sampling, and robust training (AdamW, …

Causal Discovery In Photospheric Magnetic Field Time Series For Interpretable Solar Flare Prediction, Nathan W. Nelson

All Graduate Theses and Dissertations, Fall 2023 to Present

Solar flares are capable of damaging many valuable resources, including satellites, power grids, and even human lives. Being able to predict solar flares can allow for proactive measures to be taken that can prevent that damage. Many new deep learning methods for predicting solar flares have shown promise in this task, but the decisions they make are harder to explain to humans. This makes understanding why these models make mistakes difficult, which in turn makes fixing and maintaining them more challenging. We test a recent deep learning method that helps discover relationships between different measurements of the Sun as they …

Causal Discovery Methods For Single Cell Rna-Seq Data, Melanie Lambert

All Dissertations

The advancement of single cell RNA sequencing (scRNA-seq) has enabled the study of causal relationships between genes at single cell resolution. Although many causal discovery methods have been applied to scRNA-seq perturbation data, they are not well-suited to capture the characteristics of scRNA-seq data. The overall goal of this dissertation is to enhance researchers' ability to gain insight into genetic relationships.

The scRNA-seq data is high-dimensional, typically containing thousands of genes, and is sparse and zero-inflated due to dropout events, as well as noisy and subject to biological variability. Traditional causal discovery methods, such as constraint or score-based approaches, do …

Comparison Of Log-Based Coal Structure Identification Methods For Structurally Complex Areas: A Case Study Of The Huainan Mining Area, Zhang Zixin, Zhou Xiaozhi, Chen Xuehua

Coal Geology & Exploration

Objective CO₂ displacing coalbed methane (CBM) provides substantial environmental and energy benefits. However, differences in coal structures represent a key geological factor influencing the effectiveness of this technique. Constrained by the complex geological conditions for CBM occurrence and the impacts of multiple factors on the log responses of coal structures, traditional log-based methods for coal structure identification face challenges such as much human intervention, complex feature engineering, and poor adaptability. This study aims to achieve effective identification of complex coal structures within a single coal seam. Methods An asymmetric convolution kernel-based convolutional neural network (CNN) model for coal structure …

Scalefusion: Hierarchical Feature Aggregation For Unified Multiscale Object Detection, Ayşenur Yaylaci, Berru Kaya, Mehmet Kiliçarslan

Turkish Journal of Electrical Engineering and Computer Sciences

Detecting objects across a wide range of scales, particularly small ones, remains a significant challenge in computer vision. Existing methods often improve small object detection at the cost of performance on larger objects or introduce significant computational overhead through external techniques like image slicing. This paper introduces ScaleFusion, a novel, unified, end-to-end object detection architecture designed to provide robust performance across all scales within a single model. The core of our approach is a hierarchical feature aggregation strategy structured like a tree. ScaleFusion processes an image by running a shared backbone network only on fine-grained patches at the lowest level …

Urban Carbon Emission Early Warning Research Based On The Dpsir Framework And Deep Learning, Xiaochun Zhao, Lingyang Xu, Ying Zhou

Journal of Scientific Information Research

[Purpose/significance] In line with the national requirements for building a carbon emission early warning mechanism, conducting the urban carbon emission early warning research is of great significance for achieving the “dual-carbon” goals. [Method/process] This paper selected 16 prefecture-level cities in Anhui Province as research samples. A carbon emission early warning indicator system was constructed based on the DPSIR framework. By using data from urban statistical yearbooks, the LSTM model was employed with parameter optimization via genetic algorithms to forecast various early warning indicators for 2024-2025.On this basis, a combined subjective-objective weighting method was then applied to calculate the urban carbon …

Statistical Machine Learning For Calibrating Climate Model Wind Fields., Augustine Ouru

Graduate Theses and Dissertations

This thesis investigates statistical calibration methods for wind data from Global Climate Models (GCMs), with emphasis on wind speed and wind direction. The analysis uses 30 years of daily observed and GCM wind records from 9 spatial locations in the South East USA. Wind is represented in polar coordinates, but also often modeled in Cartesian coordinates. Four calibration models are compared: a marginal wind speed model, an independent Cartesian model. The calibration models are based on semi-parametric quantile regression, a flexible conditional density estimator fitted using neural networks. The neural network objective functions are optimized using maximum likelihood estimation (MLE) …

Survey On Learning-Based Dynamic Fault Localization: From Traditional Machine Learning To Large Language Models, Chunyan Liu, Yan Lei, Huan Xie, Jinping Wang, Yue Yu, David Lo

Research Collection School Of Computing and Information Systems

Learning-based dynamic fault localization techniques play a crucial role in the field of software engineering. These techniques dynamically execute test cases to meticulously extract useful knowledge from the execution information in the program, with the aim of identifying fault locations by leveraging machine learning, deep learning, and large language models. Currently, there is already a flourishing body of research that is intensely focused on learning-based dynamic fault localization. Research literature can be categorized into two main aspects for learning-based dynamic fault localization: data-based enhancements (i.e., the datasets) and model-based enhancements (i.e., the suspiciousness algorithms). Thus, we conduct an extensive literature …

Deep Learning For Video Anomaly Detection: A Review, Peng Wu, Chengyu Pan, Yuting Yan, Guansong Pang, Qingsen Yan, Peng Wang, Yanning Zhang

Research Collection School Of Computing and Information Systems

Video anomaly detection (VAD) aims to discover behaviors or events deviating from the normality in videos. As a long-standing task in the field of computer vision, VAD has witnessed much good progress. In the era of deep learning, with the explosion of architectures of continuously growing capability and capacity, a great variety of deep learning-based methods are constantly emerging for the VAD task, greatly improving the generalization ability of detection algorithms and broadening the application scenarios. Therefore, such a multitude of methods and a large body of literature make a comprehensive survey a pressing necessity. In this article, we present …

Key Technologies For Artificial Intelligence-Based Inversion Of The Borehole Transient Electromagnetic Method In Coal Mining Areas, Fan Tao, Hao Yue, Zhang Peng, Li Ping, Zhao Zhao, Zhao Rui, Liu Qiang, Yan Junsheng, Song Xianzhe, Liu Borui, Chen Changyuan, Li Bo

Coal Geology & Exploration

Background Using directional boreholes, the borehole transient electromagnetic (TEM) method enables near-field excitation and the simultaneous reception of three components within a borehole. This method can effectively avoid interference from ferromagnetic materials in roadways, thereby significantly enhancing the detection accuracy and range. Therefore, this method is widely applied to the detection of underground concealed water hazards in coal mines. However, there is a conflict between the accuracy and efficiency of current inversion methods for borehole TEM data, and existing technologies are insufficient to simultaneously meet the demands for high efficiency and high precision in the detection of concealed water hazards. …

Seeing Around Corners: An Indirect Light Transport Decomposition Framework For Fusing Physics And Learned Priors, Fadlullah Raji

USF Tampa Graduate Theses and Dissertations

Non-line-of-sight (NLOS) imaging, or seeing around corners, is the ability to recover information about objects hidden from direct view and remains one of the most compelling challenges in computational imaging. Existing approaches fall broadly into two categories: active and passive. Active time-resolved methods achieve impressive performance but require specialized pulsed laser hardware and ultrafast detectors, making them expensive and limited by slow measurement acquisition. Passive NLOS methods, which instead exploit the subtle soft shadows (penumbrae) cast by a hidden scene onto a visible matte surface, have recently emerged as a practical alternative. However, existing passive approaches are largely limited to …

Risk-Aware Deep Reinforcement Learning In Portfolio Optimization, Shimin Zhang

USF Tampa Graduate Theses and Dissertations

Portfolio optimization is important in investment, with the purpose of maximizing profit whilemanaging risk. The traditional portfolio optimization method, Modern Portfolio Theory, is based on the assumption that returns are normally distributed, which is always violated in practical financial market settings. In addition, traditional approaches to portfolio optimization are mostly designed as static optimizations and fail to account for the nonlinearity and non-stationarity of contemporary financial markets. Recent developments in deep reinforcement learning have attracted considerable attention for treating portfolio optimization as a sequential decision-making process. Traditional portfolio optimization methods usually evaluate risk based on the variance of the returns. …

Learning Trajectories Of Online Batch Selection Methods, Luke Green

Theses and Dissertations

Modern deep neural networks achieve strong performance on large-scale datasets, but often require substantial training time. Online batch selection methods seek to reduce this cost by updating models on informative subsets of each batch rather than on all available examples. Recently introduced methods leverage teacher models and report substantial speedups, particularly in noisy-label settings. However, comparisons are often based on the number of epochs required to reach a target test accuracy, a coarse metric that is sensitive to implementation details and may obscure important differences in learning dynamics. In this thesis, we implement several online batch selection methods in a …