����Ӱ��

Harsh Dubey, ����Ӱ��

Video language models can now generate fluent, long-form paragraph descriptions for videos, but evaluation has not kept pace. Current video caption benchmarks largely focus on short captions and temporally redundant clips, and they rarely provide full-clip paragraph references or include audio/dialogue-grounded content, making them unsuitable for evaluating long-form descriptions where factual coverage and event order matter. Moreover, the same video can be described in multiple valid ways — through paraphrase, different detail levels and minor local reorder — while still preserving the underlying facts; conversely, omissions, hallucinated additions, entity/event corruption or chronology/causality breaks should be penalized. To address these gaps, we introduce VCB, a benchmark of 2,000 standardized 120-second movie-domain YouTube clips released as URLs/IDs with timestamped paragraph references. VCB includes two complementary slices: VCB-V (1k) selected for high visual nonredundancy using von Neumann entropy over frame- and subclip-level Gram spectra and VCB-AV (1k) that additionally prioritizes dialogue-dense clips using ASR-based coverage and density. We also propose VCS, a long-form evaluation score designed to be tolerant to descriptive variability while remaining sensitive to factual and ordering errors; VCS combines global and local semantic alignment with order-aware checks that tolerate limited local reorder but penalize major chronology disruptions. We further provide VCB-Suites for metric validation, including FOIL for controlled valid vs. invalid rewrites and EVAL for human preference alignment via balanced pairwise comparisons and Elo ranking. On a 100-clip subset, VCS achieves near-perfect separability on FOIL (AUC 0.97–0.99 with success@0.90 ≥ 0.95 across embedding backbones), substantially outperforming n-gram metrics (AUC 0.61–0.73) and cosine embedding baselines (AUC 0.79–0.83). In addition, VCS rankings correlate more strongly with human Elo rankings (Spearman ρ 0.78–0.83; Kendall τ 0.59–0.63) than n-gram metrics (ρ 0.30–0.42) and cosine baselines (ρ 0.58–0.66), supporting VCB as a stress test for long-form evaluation and VCS as a more reliable metric for paragraph-level video description quality.

Mukhtiar Ali, ����Ӱ��

Wearable stress detection using physiological signals has shown promising results, with models reporting accuracies exceeding 89% on benchmark datasets like WESAD. However, these results are typically obtained using window-shuffle cross-validation, which randomly splits individual time windows across training and test sets. This evaluation protocol leaks subject-specific information, as adjacent windows from the same subject appear in both splits, leading to inflated performance estimates.

We demonstrate that under strict Leave-One-Subject-Out (LOSO) evaluation, where all data from a test subject is held out during training, accuracy drops dramatically to 55% leading to a 34% generalization gap. This gap reveals that existing models memorize subject-specific patterns rather than learning generalizable stress indicators.

To address this challenge, we propose two complementary innovations. First, we introduce a hierarchical temporal model that processes sequences of consecutive windows through a GRU network, capturing stress dynamics that unfold over minutes rather than seconds. Second, we leverage explainable AI (XAI) techniques based on information-theoretic analysis to derive coinformation priors that guide multimodal fusion. These priors quantify unique information contributed by each modality and synergistic interactions between modality pairs, enabling the model to weight sensor inputs based on their predictive value rather than learning arbitrary fusion weights.

Our approach is evaluated on the WESAD dataset comprising 15 subjects with six physiological modalities from chest and wrist sensors. The hierarchical temporal context improves LOSO accuracy from 54.7% to 67.5%, and incorporating XAI-derived priors further boosts performance to 74.2% leading to a 20% improvement over baseline. Additionally, cross-subject variance is reduced, indicating more consistent generalization across individuals.

Ujjwal Pudasaini, ����Ӱ��

This paper addresses the federated unlearning problem in the smart agriculture system, where the contribution of the compromised field device must be removed from the global model under communication constraints. A three-phase framework, named FedSCAN, is proposed. FedSCAN identifies critical layers based on parameter sensitivity, classifies active neurons within these layers through relative weight change analysis and performs unlearning via sparse low-rank adaptation to active neurons while freezing base model parameters. The sparse adaptation strategy confines computational overhead to a compact parameter subset, enabling bandwidth-constrained edge devices to participate without becoming stragglers. Experimental results on three datasets with multiple architectures demonstrate that FedSCAN achieves 4.2-times to 426.3-times communication cost reduction compared to full retraining while maintaining remaining accuracy within 3% of the retraining baseline and effectively removing malicious influence.

Sohum Mallik, North Dakota State University

Hepatocellular carcinoma (HCC), the most common form of primary liver cancer, remains a leading cause of cancer-related mortality worldwide. Limited therapeutic options and high recurrence rates highlight the need for more efficient strategies to identify promising anticancer agents. Computational modeling offers an effective approach to support early-stage drug discovery by prioritizing compounds prior to resource-intensive experimental testing.

In this study, we evaluated whether an interpretable quantitative structure-activity relationship (QSAR) model could reliably predict anticancer activity of small molecules in the Huh-7 hepatocellular carcinoma cell line. A dataset of compounds with experimentally reported IC₅₀ values in Huh-7 cells was compiled from the literature. Molecular structures were prepared and optimized prior to descriptor calculation, and QSAR models were developed using multiple linear regression within the QSARINS framework.

Model robustness and predictive performance were assessed using leave-one-out and leave-many-out cross-validation, an external test set, Y-scrambling and applicability domain analysis based on the Williams plot. The final model demonstrated strong statistical performance (R²_train = 0.896, Q²_LOO = 0.848, Q²_LMO = 0.826, R²_ext = 0.826) with low prediction errors for both training (RMSE = 0.108) and external (RMSE = 0.128) sets. The external concordance correlation coefficient (CCC) was 0.888, and Y-scrambling confirmed the absence of chance correlation.

Descriptor analysis identified H0v and CATS2D_04_LL as key contributors to anticancer activity, providing interpretable insight into the structural features influencing potency. Overall, these results suggest that the developed QSAR model is a reliable tool for preliminary virtual screening of organic compounds with potential anticancer activity in HCC, supporting more efficient prioritization in early drug discovery.

Mohammad Johurul Islam, ����Ӱ��

Cascading failures can escalate from an initial transmission line outage into widespread blackouts. This poster presents a deep learning framework that learns component interactions from utility outage data. A Long Short-Term Memory (LSTM) sequence model with multihead attention captures outage propagation dependencies across stages and the learned relationships are aggregated into a global interaction matrix for ranking critical transmission lines. In experiments, the framework achieves 70% Top-20 recall of benchmark critical lines under the combined (C1 ∪ C2) criterion. Future work includes transformer/LLM-inspired long-range models, CCDF-based validation using simulated cascades, mitigation-impact evaluation and a desktop interactive tool.

Zain ul Abideen Usmani, ����Ӱ��

Accurate estimation of foliar nitrogen concentration (N%) is essential for precision agriculture, as nitrogen strongly influences crop growth, yield formation and fertilizer use efficiency. Conventional laboratory-based measurements are time-consuming, costly and impractical for large-scale monitoring. Hyperspectral remote sensing offers a powerful alternative; however, purely data-driven models often lack physical interpretability and show reduced robustness across phenological stages due to stage-dependent changes in leaf optical properties. This study presents a hybrid deep learning framework for predicting leaf nitrogen concentration from hyperspectral reflectance and radiatively derived biochemical parameters using a Stage-Aware Feature-Wise Linear Modulation (FiLM) Multi-Layer Perceptron (MLP). Hyperspectral measurements in the 397.5-1,010.9 nm range were collected from corn (Zea mays L.) leaves at vegetative growth stages V5 to V8 under three nitrogen treatments (0N, 80N, and 120N). Ground-truth nitrogen content was obtained through standard laboratory chemical analysis. A radiative transfer model was employed to retrieve physically meaningful biochemical features and to enforce consistency between spectral responses and leaf optical properties. Growth stage information was incorporated through FiLM conditioning, enabling the network to adaptively modulate its internal representations for each phenological stage while preserving shared spectral and biophysical relationships. All variables were standardized and used to train the model with stable optimization. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE) and mean absolute error (MAE). Compared with baseline methods, including Support Vector Regression, Partial Least Squares Regression and Random Forest, which achieved R² values ranging from 0.36 to 0.81 with higher prediction errors, the proposed Stage-FiLM MLP achieved superior performance with an R² of 0.85, RMSE of 0.21 and MAE of 0.13. These results demonstrate that integrating radiative transfer constraints with stage-conditioned deep learning enhances accuracy, cross-stage predictive stability and physical interpretability, making the approach suitable for large-scale precision nitrogen management in maize.

Zihao Ding, ����Ӱ��

Knowledge variety and knowledge forgetting are two major issues in sustaining collaborative learning within heterogeneous vehicle networks. These issues become especially severe when vehicles possess varying sensing capabilities, computational resources and domain expertise, leading to fragmented learning and unstable knowledge retention over time. To address these challenges, we propose a dual-level game-theoretic approach. We first formulate a new metric, Utility-of-Information (UoI), to characterize the features of knowledge learning, retention and consolidation. Based on this metric, we design a game-theoretic dual-level approach, which comprises a lower-level coalition formation game where vehicles self-organize into "teacher-student'' coalitions based on their UoI profiles, and an upper-level UAV resource allocation game where vehicle coalitions compete for limited communication resources. To optimize both levels of the game, we design a unified reinforcement learning-based framework that enables adaptive searching for optimization under dynamic network conditions. Experimental results demonstrate that our approach effectively addresses knowledge variety and significantly mitigates the effects of knowledge forgetting in UAV-assisted heterogeneous vehicle networks.

Dongyoun Kim, ����Ӱ��

Vision-language models (VLMs) like CLIP have achieved remarkable success in image-text alignment, but they are optimized for image-level representations. Existing region-level methods focus primarily on single-object regions.

We propose CLIP-MORE (Multi-Object Region Embeddings), a lightweight framework that distills CLIP's global knowledge into efficient multiobject region representations. Our approach addresses three key challenges: (1) the global-local representation mismatch in VLMs, (2) the single-object focus of existing methods and (3) limited supervision signals for distillation.

To bridge global and local representations, we explore a multiobjective distillation loss combining Mean Absolute Error (MAE), contrastive learning (SimCLR) and Soft Contrastive Learning (SoftCLR). SoftCLR captures partial semantic overlap between region-text pairs using continuous similarity weights rather than hard binary labels.

We train on 2.5M multiobject regions extracted from COCO+LVIS (118K training images). Our lightweight 17.7M parameter encoder achieves significant improvements over the 86.2M parameter CLIP: +83% on Text-to-Image R@1 (0.286 → 0.524) and +77% on Image-to-Text R@1 (0.282 → 0.498). Adding SoftCLR to SimCLR alone improves T2I R@1 by +11% (0.472 → 0.524).

Our results demonstrate that effective multiobject region representations can be achieved with significantly smaller models through carefully designed distillation objectives.

Preethi Amasa, ����Ӱ��

Video anomaly detection (VAD) is difficult because anomalies are highly varied and often defined by scene context rather than explicit actions. Retrieval based methods such as Flashback provide training free and interpretable detection by matching video embeddings to textual memories, but their performance varies across benchmarks. While Flashback performs well on action centric datasets such as UCF-Crime and XD-Violence, it degrades on ShanghaiTech, where anomalies are subtle and arise from abnormal motion patterns or spatial usage.

This poster analyzes why semantic retrieval fails on ShanghaiTech and proposes targeted modifications that preserve Flashback’s retrieval framework while improving sensitivity to contextual deviations. First, we redesign prompt and memory construction by removing explicit action category tokens from anomalous captions, maintaining the repulsive prompt structure (“Normal scene:” vs. “Anomalous scene:”), and building class separated text memory using frozen ImageBind text embeddings. This reduces semantic bias that can dominate similarity scores and lead to false positives in context-driven scenarios.

Second, we introduce lightweight temporal modeling by aggregating multiframe appearance embeddings and incorporating a motion difference cue, enabling the representation to emphasize dynamic irregularities rather than static semantics. Finally, we integrate a Mixture-of-Experts Low-Rank Adaptation (MoELoRA) module into the frozen ImageBind vision encoder using lightweight LoRA adapters and a learned routing network that selects expert pathways for each video segment. Only the router and LoRA parameters are trained on normal ShanghaiTech clips using a retrieval-based margin objective, allowing efficient and context-adaptive refinement of video embeddings, while preserving the backbone, retrieval pipeline and text-based explainability.

Preliminary results on ShanghaiTech show that temporal aggregation and prompt/memory tuning outperform a Flashback baseline, and ongoing experiments evaluate the contribution of MoELoRA. This work supports forthcoming research on scalable, interpretable and context-adaptive retrieval for video anomaly detection.

RKBM Rizmi, ����Ӱ��

This study establishes a systematic benchmarking framework for evaluating data-driven predictors in structural dynamics, utilizing validated finite element data of a cantilever beam undergoing damped free vibration. The methodology integrates classical Euler-Bernoulli beam theory to provide an analytical ground truth, ensuring the fidelity of transient structural simulations performed in ANSYS. Using this high-fidelity data, four distinct modeling architectures — Delay-Embedded Dynamic Mode Decomposition (DMD), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks and Physics-Informed Radial Basis Function Neural Networks (PI-RBFNN) — are rigorously compared for predictive accuracy.

Quantitative analysis using Root Mean Square Error (RMSE) reveals that DMD achieves superior performance, outperforming the neural network-based architectures in both accuracy and computational efficiency for free vibration responses. These findings demonstrate that DMD is a robust surrogate for linear structural dynamics and establish a foundation for extending this benchmarking framework to more complex, nonlinear dynamical systems.

Muhammad Roman, ����Ӱ��

Accurate monitoring of nitrate and nitrite concentrations in water is essential for sustainable agriculture, safeguarding public health and protecting aquatic ecosystems from nutrient pollution. Traditional methods for detecting nitrate and nitrite in water samples are precise but suffer from high costs, complexity and delays, which limit their practicality for frequent, on-site testing. This research proposes deep learning-based computer vision techniques to classify nitrate and nitrite concentrations using colorimetric test strip images. An RGB IMX219 camera was used to acquire colorimetric test strip images under standardized, controlled illumination conditions to ensure consistent image quality. A total of 1,938 nitrate images and 1,190 nitrite images were collected before augmentation. After preprocessing and training-only data augmentation, both classical machine learning baselines based on hand-crafted color and texture features and deep learning models — including a Multilayer Perceptron (MLP) and Convolutional Neural Networks (AlexNet, VGG16, ResNet18 and GoogLeNet) — were trained and evaluated using an independent test set and stratified five-fold cross-validation. For nitrate classification, ResNet18 and GoogLeNet achieved near-perfect 100% test accuracy, with mean cross-validation accuracy of 99.97% ± 0.04%, substantially outperforming classical baseline models based on hand-crafted color and texture features, which achieved at most 83.5% test accuracy. For nitrite classification, GoogLeNet achieved the strongest overall performance, with a test accuracy of 97.48% and a five-fold cross-validation accuracy of 95.22% ± 1.17%, substantially outperforming the best classical baseline model, which achieved a maximum test accuracy of 83.19%. These results demonstrate that deep CNN-based feature learning provides a significant performance advantage over simpler methods under controlled imaging conditions, supporting the suitability of the proposed system for rapid, image-based water quality assessment and motivating future evaluation under broader real-world deployment scenarios.

Jingong Chin, ����Ӱ��

1. Bridge the Gap in Identity Unlearning Capability of Existing ReID Models
   1. Current person ReID models struggle to efficiently forget specified identities posttraining. Traditional methods require reoptimization, which is time-consuming and prone to damaging the retrieval structure of nontarget identities, failing to meet dual demands of fast deployment and performance preservation.
2. Meet the Inference-Time Unlearning Requirement Without Retraining
   1. Practical scenarios frequently demand responses to identity data deletion requests. Existing schemes rely on training-phase adjustments and cannot realize unlearning by simply removing components during inference, making them incompatible with privacy compliance and flexible deployment needs.
3. Resolve the Coupling Issue of Target and Nontarget Feature Representations
   1. Traditional unlearning methods tend to contaminate the base feature space, degrading the retrieval performance of nontarget identities. A responsibility migration mechanism is needed to route target identity features to an independent pathway, decoupling target forgetting from nontarget performance maintenance.

Benjamin Van De Sande, Augsburg University

High-frequency cryptocurrency markets evolve rapidly, and their microstructure generates large volumes of noisy, heterogeneous data. Traditional supervised learning methods perform poorly in this setting due to limited label availability and rapidly shifting data distributions.

As a result, adaptive and unsupervised methods for extracting meaningful structure from real-time financial data streams are gaining increasing prominence. This project presents an end-to-end data acquisition system and exploratory modeling pipeline for high-frequency Bitcoin market data.

The first stage of the project involves developing a custom hardware-based scraper deployed on a Raspberry Pi 5 to establish a functional hardware-to-model pipeline that streams raw subsecond Bitcoin market microstructure data from the Kraken exchange into a unified data lake. Within this data lake, the incoming streams are systematically cleaned, synchronized and aggregated into consistent, analysis-ready one-second frames that are further augmented with learnable features to support downstream modeling and analysis. Building the dataset directly from raw exchange feeds provides full control over data quality, enables experimentation with streaming and label-free learning methods and ensures transparency and reproducibility across all downstream analyses.

In the second stage, the project explores unsupervised representation learning and online regime detection through two prototype modeling approaches: a baseline clustering-based method and a learned embedding method paired with a streaming anomaly detector. Preliminary experiments surfaced structural patterns in unlabeled microstructure data while offering limited direct predictive capabilities. Embedding-based anomaly detection highlighted moments of unusual market activity, while clustering analyses revealed interpretable relationships among engineered variables. These promising results show that the pipeline produces coherent signals and supports an exploratory study of evolving market conditions.

This work provides a foundation for future research in label-free financial analytics, real-time decision systems and adaptive modeling of nonstationary data.

Zhongyu Xie, ����Ӱ��

Procedural Content Generation (PCG) is a core technique for creating large-scale virtual environments, but generating high-quality and varied 3D content remains challenging, especially under limited data settings. Recent diffusion-based approaches, such as Diffusion Craft, have demonstrated strong fidelity in Minecraft level generation from a single example. However, these methods typically require hundreds to thousands of denoising steps, resulting in high sampling latency and limited variety among generated levels.

In this work, we explore Flow Matching (FM) as an alternative generative paradigm for procedural content generation in Minecraft. Instead of learning a long stochastic diffusion trajectory, Flow Matching learns a direct velocity field that defines a simple probability path between noise and data representations. This formulation allows efficient generation using only a small number of integration steps. To enable continuous modeling, we adopt Block2Vec to embed discrete Minecraft blocks into a dense vector space and train a 3D U-Net to predict the velocity field along the flow path.

We evaluate our approach across multiple Minecraft biomes, including ruins, desert, plains, swamp and village environments. Generation quality is measured using Tile-Pattern KL Divergence (TPKL-Div) to assess local structural fidelity, while variety is quantified using pairwise Levenshtein distance between generated samples. Sampling efficiency is evaluated using Latency@85%, defined as the time required to reach 85% structural accuracy during generation.

Our results show that while Flow Matching does not consistently outperform Diffusion Craft in terms of fidelity, it achieves significantly lower sampling latency and higher intersample variety across most biomes. These findings suggest that Flow Matching offers a promising trade-off between quality, speed and variety, making it a compelling direction for efficient procedural content generation in large-scale 3D environments.

Lalit Kharel, University of Nebraska

Survival prediction for lung adenocarcinoma (LUAD) remains challenging due to heavy censoring in clinical datasets. To address data sparsity, ensemble machine learning has shown promise, while most deep learning models sacrifice interpretability and risk overfitting or unstable performance. This work aims to provide an interpretable “glass-box” framework that achieves both high interpretability and improved survival prediction in patients with LUAD. To select the most discriminative subset of features, penalized CoxNet was applied to LUAD cases from The Cancer Genome Atlas (TCGA). The resulting feature subset was then used to train four survival models: Random Survival Forest (RSF), Gradient Boosting Survival Analysis (GBS), Extreme Gradient Boosting (XGB) and DeepSurv Neural Network. The model predictions from out-of-fold samples were combined using a generalized additive model (GAM) meta-learner, which provides transparent weighting for the contribution of each model to the ensemble. Cross-validated ablation analysis demonstrated that RSF, GBS and DeepSurv consistently contributed positively to ensemble performance, while XGB provided no measurable benefit. Based on these findings, we constructed an optimized three-model ensemble (RSF + GBS + DeepSurv). This model achieved a concordance index (C-index) of 0.671 (64% DeepSurv, 21% GBS, and 15% RSF), outperforming the four individual models while remaining fully interpretable. Kaplan-Meier analysis was used to evaluate model-derived risk stratification, demonstrating strong clinical stratification, with high-risk patients exhibiting significantly shorter survival compared to the low-risk group (log-rank p = 3.06 × 10−3). Although the medium-risk group did not reach median survival during follow-up, survival curves showed a clear ordinal separation between risk strata. These results demonstrate that combining biologically grounded feature selection with interpretable ensemble learning yields improved predictive accuracy and clinically meaningful risk stratification for LUAD, supporting the use of AI models in precision oncology.

Dhe Yeong Tchalla, ����Ӱ��

Accurate traffic forecasting is fundamental for intelligent transportation systems, directly influencing congestion management, safety and emissions. A central challenge is the nonstationary nature of traffic flow, where relationships between sensors shift dynamically due to incidents, demand changes and propagation waves. Many existing graph-based models rely on static graphs or computationally heavy dynamic mechanisms, limiting their suitability for real-time deployment. We propose ST-Hybrid, a spatiotemporal model that incorporates a lightweight state-conditioned dynamic graph learner. The module updates connectivity in real-time using adaptive node embeddings and sparse top-𝑘 neighborhood selection, offering a practical balance between flexibility and efficiency. On PeMSD8, it obtains a mean absolute error (MAE) of15.19 and root mean square error (RMSE) of 24.18, ranking second among recent dynamic-graph approaches. It maintains competitive accuracy on the larger PeMSD4 network (MAE 20.16, RMSE 32.17) and demonstrates robust performance on PeMSD3 (MAE 15.58,RMSE 25.93) and PeMSD7 (MAE 21.58, RMSE 34.36). Sensor-level analyses reveal that most residual error is concentrated in a small subset of volatile sensors, suggesting that targeted refinement may yield greater improvements than further global architectural complexity. Importantly, ST-Hybrid maintains low inference latency (approximately 15 ms), demonstrating its suitability for large-scale, real-time traffic forecasting applications.

Joy Okolo, ����Ӱ��

Federated learning (FL) faces significant challenges in heterogeneous mobile edge environments, where disparities in device capabilities create stragglers that delay training and force trade-offs between all-encompassing and efficiency. Existing methods either exclude slower clients, reducing fairness and generalization, or prolong training by waiting for them, wasting resources on faster devices.

We propose Client-Oriented Adaptive Model Pruning (CLAMP), an adaptive framework that dynamically adjusts the number of trainable neural network layers per client based on real-time computational performance and latency feedback. CLAMP employs client-specific thresholds with validation-gated depth adjustment and stability control to prevent oscillations and ensure convergence. Unlike static partial training or predefined heterogeneous architectures, CLAMP continuously adapts to runtime conditions without complex model design or heavy communication overhead. Extensive experiments on benchmark datasets show that CLAMP maintains competitive accuracy while significantly improving all-encompassing and resource utilization. It increases slow-device participation, reduces total training time and achieves stable convergence, making CLAMP a practical and deployable solution for heterogeneous federated learning where both model quality and fair participation are critical.

Synthia Ferdouse, ����Ӱ��

Modern engineering structures such as aircraft panels, bridges and industrial components constantly experience loads, vibrations and environmental effects that can create small defects, which may grow into serious failures if not detected in time. Structural Health Monitoring (SHM) helps address this problem by continuously assessing the condition of structures, improving safety while reducing inspection time and maintenance costs. Ultrasonic guided waves, together with surface-mounted piezoelectric sensors, provide a sensitive and nondestructive way to evaluate the internal condition of structures. However, interpreting these complex signals usually requires expert knowledge and extensive signal processing, making real-time and large-scale applications challenging.

To address this limitation, this study presents an automated damage detection framework that combines physics-based simulation and machine learning. An aluminum plate was modeled in ANSYS with two PZT sensors, where one sensor introduced random excitation and the other recorded the resulting wave propagation responses. Baseline signals from the healthy plate and signals from plates containing four different notch-type defects were collected, forming datasets representing intact and damaged structural conditions under varying noise levels. A 1D Convolutional Neural Network (1D-CNN) was trained to learn discriminative features directly from raw wave signals without manual feature extraction.

The results demonstrate that the trained model consistently and accurately distinguishes between healthy and damaged wave responses when evaluated using previously unseen data, with predictions closely aligned with the simulated defect scenarios, highlighting the model’s reliability, strong generalization capability and effectiveness for data-driven structural damage identification.

Mohammad Yeasin, ����Ӱ��

Regional deposition of inhaled aerosols in the human respiratory tract plays a central role in airborne infectious disease transmission and the effectiveness of inhaled drug delivery. While airway anatomy is highly variable, particle deposition is fundamentally governed by physical mechanisms including particle inertia, pressure-driven acceleration and regional flow partitioning. A key unresolved challenge is how to extract generalizable, low-dimensional structure from high-fidelity simulation data while preserving physical interpretability across anatomically distinct airways.

In this study, we combine physics-based CFD simulations with artificial intelligence to learn a constrained, low-dimensional representation of regional particle deposition. Ten anatomically realistic human respiratory geometries are analyzed under five pressure-driven inhalation conditions. Each geometry is segmented into three functional regions — nasal cavity, pharynx and lower airway — and Lagrangian particle tracking is performed for particle diameters ranging from one to 20 µm in 0.5 µm increments. For each monodisperse case, regional deposition fractions are computed and embedded into a ternary phase space, enforcing mass conservation and encoding the competitive nature of regional particle capture.

Despite substantial geometric variability, deposition outcomes across all anatomies organize into smooth, structured trajectories within ternary space as particle size and inhalation pressure vary. These trajectories reveal consistent regime transitions between nasal-dominated, pharyngeal-buffered and lower-airway-dominated deposition. To characterize this structure, AI-based manifold learning and regression models are employed to reconstruct continuous deposition surfaces within the physically admissible ternary domain. The learned representations capture dominant physics-driven trends while filtering geometry-specific variability.

Rather than serving as a black-box predictor, AI is used here as a physics-informed representation and interpolation tool, enabling rapid mapping of deposition regimes as functions of particle size and flow conditions. This framework demonstrates how constrained machine learning can extract low-dimensional physical structure from complex biomedical simulations, providing scalable insights for airborne infection risk assessment and inhaled therapeutic design.

Sjairua Sehgal, ����Ӱ��

Crop lodging can substantially reduce grain yield and degrade grain quality; therefore, rapid and accurate assessment of crop lodging is a critical step in plant breeding selection, and it directly affects variety release decisions. Lodging information is also essential for field management and timely decision-making, including harvest planning, risk mitigation and postevent damage assessment. Traditional field scouting is labor-intensive, time-sensitive and often subjective, making it difficult to scale across large breeding trials or commercial fields and prone to missing spatially heterogeneous lodging patterns. Uncrewed Aircraft Systems (UAS)-based remote sensing surveys can rapidly cover nurseries at high spatial resolution and can be repeated after wind or rain events, enabling consistent, plot-level lodging scoring and explicit decision support. 

In this study, we integrated UAS multispectral imagery with deep learning (DL) to quantify lodging severity in wheat breeding trials in South Dakota (July 2025). Orthomosaicked multispectral bands were combined with UAS-derived canopy height information to capture both spectral responses and structural signatures associated with lodging. DL methods such as Convolutional neural networks (CNNs) with data augmentation were trained to classify lodging intensity into four categories: no lodging, mild lodging, medium lodging and severe lodging.

The results show overall good performance across severity levels, with reasonable agreement with ground-truth observations, while still exhibiting some confusion between adjacent classes. These findings highlight the practical potential of DL-enabled UAS remote sensing for scalable lodging assessment and phenotyping, while underscoring the need for broader multisite, multiyear evaluation and further refinement to improve robustness under variable field conditions.

Sheetalpreet Kaur, ����Ӱ��

Candidate orphan genes (COGs) lack homologues outside given taxon, have ambiguous evolutionary origins and are often associated with lineage-specific traits, including pathogenicity and ecological adaptation. Prompted by unresolved nature of COGs, we recently reanalyzed COGs reported in 2023 from TRG database. After reanalysis, we determined COG numbers were substantially overestimated and only a small subset remained lineage-specific. From 10 bacterial species, we identified only 10 high-confidence COGs predicted to encode bona fide proteins. Building on our previous work, we address whether these loci represent true genes or annotation artifacts, remain taxonomically restricted in expanded environmental datasets, encode bona fide proteins and perform defined cellular functions. As proof of concept, we examined one COG present in B. subtilis and confirmed its presence in environmental samples using PCR. To determine the function of this gene, we are employing multiple complementary strategies, including genome mutagenesis/deletion studies and promoter-reporter assays. In parallel, AI-based protein structure predictions will be used to identify similarities to known protein folds, and gene-centered pull-down assays will be explored to identify potential protein interaction partners. In addition, the genomic context of a gene can provide important clues to its function, as bacterial genes are often organized into operons. Operon context often enables functional assignment of unknown genes, but COGs are rarely found in recognizable operons. We hypothesize that many COGs were originally part of operons but became isolated over evolutionary time. To investigate gene neighborhood conservation in environmental samples, we will use a hybridization-capture approach in which tagged DNA probes bound to streptavidin beads capture the target gene and surrounding genomic DNA. Sequencing captured fragments will enable analysis of neighboring genomic regions and comparison of gene neighborhoods and operon structures across organisms. Together, these complementary approaches aim to determine whether COGs encode functional proteins with roles in their bacterial host.

Beining Wu, ����Ӱ��

Reconfigurable Intelligent Surface (RIS)-assisted Unmanned Aerial Vehicle (UAV) communications have been realized as essential to space-air-group system integration in the 6G technology landscape. Trajectory planning plays a crucial role in RIS-assisted UAV communications to face the challenges of UAV's limited power capacities and dynamic wireless channels. Existing solutions assume complete channel state information, focus on single-rotor UAVs and rely heavily on time-consuming training processes for machine learning; thus, they lack applicability to deal with highly dynamic real-world scenarios. To fill these research gaps, we aim to characterize RIS-assisted UAV communications and design responsive and accurate UAV trajectory planning algorithms in this paper. We first develop a communication model with incomplete information and an energy consumption model for quadrotor UAVs. We then formulate UAV trajectory planning as an optimization problem to minimize UAV's energy consumption while maintaining communication throughput. To solve this problem, we design an acceleration framework, {\sl FedX}, for reinforcement learning (RL) solvers and present two fast trajectory planning algorithms, FedSAC and FedPPO, as instantiations of the {\sl FedX} framework. Our evaluation results indicate that the proposed framework is effective and efficient — more than three times faster with five agents and seven times faster with 10 agents than standard RL algorithms, making it suitable for using RL solvers within wireless networks and mobile computing environments. We also discuss and identify the pros and cons of our proposed framework.

Ruyi Lian, ����Ӱ��

Given a single input RGB image, the instance-level 6DoF object pose estimator recovers rotation and translation of a rigid object with respect to a calibrated camera. The pose estimator is crucial in numerous real-world applications, including robot manipulation, autonomous driving, augmented reality, etc. To increase the robustness under various imaging conditions, most existing methods first generate correspondences between 2D image pixels and 3D object points, and then regress the pose via any available Perspective-n-Point (PnP) solver.

However, unreliable localization results of invisible key points degrade the quality of correspondences. In this work, we address this issue by localizing the important key points in terms of visibility. Since key point visibility information is currently missing in the dataset collection process, we propose an efficient way to generate binary visibility labels from available object-level annotations, for key points of both asymmetric objects and symmetric objects. We further derive real-valued visibility-aware importance from binary labels based on the PageRank algorithm.

Our visibility-aware importance can be seamlessly integrated into existing key point-based 6DoF pose estimator to boost performance. We add a visibility-aware importance predictor before key point localization to eliminate key points with low importance. We use positional encoding to enhance the embeddings of the selected key points and adopt a two-stage training strategy to efficiently train our pose estimator. Furthermore, our method can easily adapt to a more general setting where precise 3D CAD model is unavailable.

Extensive experiments are conducted on popular pose estimation benchmarks including Linemod, Linemod-Occlusion and YCB-V, demonstrating that VAPO clearly achieves state-of-the-art performances, while improving robustness and reliability through explicit visibility reasoning.

Khalid Mahmud Labib, ����Ӱ��

The complex electrochemical behavior of lithium-ion batteries results in nonlinear dynamics and appropriate modeling of this nonlinear dynamical system is of interest for better management and control. In this work, we proposed a family of dynamic mode decomposition (DMD)-based data-driven models that do not require detailed knowledge of the composition of the battery materials but can essentially capture the nonlinear dynamics with higher computational efficiency. Only voltage and current data obtained from hybrid pulse power characterization (HPPC) tests were utilized to form the state space matrices and subsequently used for predicting the future terminal voltage at different state of charge (SoC) and aging levels. To construct the system model, 60% of the data from a single HPPC test was utilized to generate time-delay embedded snapshots, with embedding dimension ranging from 40 to 2,000. Among these, an embedding dimension of 1,810 resulted in the least residual sum of squares (RSS) error of 3.86 for the dynamic mode decomposition with control (DMDc) model and 30 for the standard DMD model. For DMDc model, delay embeddings (ranging from one to 12) were also incorporated into the input current signals. For the input matrix, an embedding dimension of six resulted in a minimum RSS error of 1.74. Furthermore, the system matrices A and B, identified from the HPPC test when the cell is in its healthy state, were held fixed and used to simulate the system dynamics for aged batteries by updating only the control input. Despite the presence of nonlinear degradation effects in later cycles, the DMDc model effectively captured key inner dynamics such as voltage dips and transient responses for subsequent charge and discharge cycles.

Subash Thapa, ����Ӱ��

Fusarium head blight (FHB), caused primarily by Fusarium graminearum, is a major fungal disease of wheat that reduces grain yield and quality and contaminates grain with mycotoxins. Accurate and scalable phenotyping of FHB severity and Fusarium-damaged kernels (FDK%) is critical for resistance breeding but remains constrained by subjective, labor-intensive visual scoring. In this study, two state-of-the-art deep learning models, YOLOv11 and YOLOv12, were implemented using oriented bounding box (OBB) annotations to automate the detection and classification of healthy and infected spikelets and kernels from high-resolution RGB images. A large-scale benchmark dataset, the SD-FHBSD and SD-FDKD, was developed containing precisely annotated spikelet- and kernel-level images collected from a controlled FHB nursery. YOLOv11 achieved superior detection and classification performance, with mAP@0.5 values of 0.937 for spikelets and 0.932 for kernels, and the fastest inference time (9.5-14.2 ms per image), supporting real-time phenotyping applications. For FHB severity estimation, YOLOv11 reached R² = 0.96 with the lowest RMSE (0.06), while for FDK%, it achieved R² = 0.92 and RMSE = 0.05, outperforming YOLOv12. Incorporation of brightness-based augmentation further improved robustness across lighting conditions. The integration of OBBs enabled more precise localization of tilted and overlapping spikelets, enhancing detection accuracy compared with conventional axis-aligned boxes. This study introduces the first OBB-based FHB dataset and demonstrates an efficient, low-cost and scalable framework for automated disease severity and kernel damage estimation, paving the way for AI-driven, real-time phenotyping and precision disease monitoring in wheat breeding programs.

Moditha Reddy Jagannath Reddy, ����Ӱ��

Cisplatin-based chemotherapy is a standard treatment for bladder cancer; however, the emergence of cisplatin resistance remains a major clinical challenge, frequently resulting in tumor recurrence and poor patient outcomes. While genetic alterations have been studied extensively, increasing evidence suggests that epigenetic mechanisms play a critical role in regulating treatment response. In particular, DNA methylation can alter chromatin accessibility, thereby influencing gene expression programs associated with drug resistance. Despite its importance, the complex relationship between DNA methylation and chromatin accessibility in cisplatin resistance is not yet fully understood.

In this study, we apply interpretable machine learning approaches to investigate epigenetic patterns associated with cisplatin resistance using large-scale DNA methylation and chromatin accessibility data. An HM450 methylation assay-derived dataset was integrated with chromatin accessibility profiles, CpG island annotations and genomic feature information. Multiple interpretable models were employed, including Association Rule Mining, Bayesian Belief Networks, Decision Tree classifiers and K-mode clustering, enabling transparent analysis of multivariate epigenetic interactions.

Our results consistently demonstrate that DNA hypermethylation is strongly associated with reduced chromatin accessibility across multiple models. Bayesian Belief Networks reveal probabilistic dependencies between methylation changes and accessibility-related genomic features, highlighting key epigenetic interactions underlying resistance. K-mode clustering identifies distinct epigenetic profiles, revealing heterogeneity among resistant samples and suggesting the presence of multiple resistance-associated subtypes. Decision Tree models further provide human-readable, threshold-based rules that distinguish hypermethylated and hypomethylated chromatin states, reinforcing the interpretability and robustness of the findings.

Overall, this work demonstrates the value of interpretable machine learning for uncovering biologically meaningful epigenetic mechanisms underlying cisplatin resistance. These insights support epigenetic biomarker discovery and contribute to the development of personalized therapeutic strategies for bladder cancer.

Nirmalya Thakur, ����Ӱ��

This work presents the ����Ӱ�� Sentiment and Engagement Index (CSEI), developed to capture nuanced variations in public sentiment and engagement on social media, particularly in response to major COVID-19-related events. Developed with various sentiment indicators, CSEI integrates features such as engagement, daily post count, compound sentiment, fine-grained sentiments (fear, surprise, joy, sadness, anger, disgust and neutral), readability, offensiveness and domain variability. Each component is systematically weighted using a multistep Principal Component Analysis (PCA)-based framework, prioritizing features according to their contributions to variance across temporal sentiment shifts. This approach dynamically adjusts component importance, enabling CSEI to precisely capture high-sensitivity shifts in public sentiment. CSEI’s development showed statistically significant correlations with its constituent features, underscoring internal consistency and sensitivity to specific sentiment dimensions. CSEI's responsiveness was validated using a dataset of 4,510,178 Reddit posts about COVID-19. The analysis focused on 15 major events, including the WHO's declaration of COVID-19 as a pandemic, the first reported cases of COVID-19 across different countries, national lockdowns, vaccine developments and crucial public health measures. Cumulative changes in CSEI revealed prominent peaks and valleys aligned with these events, indicating significant patterns in public sentiment across different phases of the pandemic. Pearson correlation analysis further confirmed a statistically significant relationship between CSEI daily fluctuations and these events, highlighting CSEI’s capacity to infer and interpret the shifts in public sentiment and engagement in response to major events related to COVID-19.

Xuan Butzin, ����Ӱ��

The accelerating spread of antibiotic-resistant pathogens has created an urgent need for novel antimicrobial strategies beyond conventional targets such as cell division, DNA replication, transcription and translation. One alternative approach is hyperactivation of bacterial proteases, which induces uncontrolled protein degradation and rapid cell death. ADEP4, a cyclic peptide antibiotic, exemplifies this mechanism by binding the ClpP protease and triggering unregulated proteolysis. Inspired by this mode of action, we explore structure-based discovery of small-molecule bacterial protease activators.

We employed in silico screening approaches combining molecular docking, structural analysis and drug-likeness and pharmacokinetic filtering, using known protease binding sites as templates and binding affinity as the primary selection criterion. Early screening efforts identified multiple chemically varied small-molecule candidates with favorable predicted interactions, supporting the feasibility of identifying nonpeptidic protease activators through computational methods. Complementary wet-lab validation experiments demonstrated that two molecules selectively killed Gram-positive bacteria, providing functional validation of this discovery pipeline.

These results motivated expansion of virtual screening to a substantially larger chemical library, enabling broader and more systematic exploration of chemical space. However, large-scale molecular docking is computationally intensive and limits scalability. To address this challenge, we are developing an artificial intelligence-based prioritization framework to augment traditional docking workflows. By preranking compounds based on learned interaction patterns, AI models can reduce the number of candidates subjected to high-cost docking calculations. Integrating AI-driven prioritization with structure-based virtual screening aims to improve the efficiency, scalability and throughput of antibiotic discovery targeting bacterial protease hyperactivation.

Jacob James, South Dakota School of Mines

Many real-world time series problems encounter change in the incoming data, whether that be degrading sensors or changes in the data being collected. The work in this poster attempts to solve the case where this change in incoming data happens at a constant and continuous rate. Simulated data was created in the form of binary classification in order to test the capabilities of the Bayesian Temporal Neural Network (BTNN). The BTNN enforces a Gaussian Process on a set of neural network parameters over time, allowing for smooth updates. To reduce memory and time complexity, two techniques were enforced and tested: A transportation matrix to transport a smaller network to the original larger network architecture, Linear Interpolation across time of the model parameters. Each method was tested with respect to accuracy and computation time, providing a solid foundation to expand to more complicated model architectures and datasets.

Raghavi Janaswamy, ����Ӱ��

The NCBI Gene Expression Omnibus (GEO) is one of the largest public functional genomics repositories, containing over 1 million samples from approximately 200,000 studies. This vast collection supports biomarker discovery, disease characterization and cross-study meta-analysis, but remains difficult to efficiently search, and integrate with external analytical systems. Each study includes a structured gene-expression matrix and unstructured textual metadata related to experimental design and sample characteristics. However, metadata quality and consistency vary widely across submissions. Terminological heterogeneity (e.g., “liver cancer” versus “hepatoma”), missing annotations and inconsistent formatting hinder systematic retrieval, integration and large-scale reanalysis. Consequently, many valuable studies remain difficult to identify and reuse.

This project develops a scalable, AI-driven framework for automated metadata curation and intelligent study retrieval within GEO. The system integrates Retrieval-Augmented Generation (RAG) with Medical Subject Headings (MeSH)-based semantic indexing and keyword search. This enables the hybrid discovery, normalization and categorization despite incomplete and inconsistent annotations. By combining large language models with biomedical ontologies and information retrieval methods, the framework extracts, standardizes and enriches study-level metadata. The system supports complex, concept-driven queries and returns explainable, context-aware results aligned with standardized MeSH terminology. Ultimately, this work aims to make GEO more accessible and interoperable, lowering barriers to data reuse and accelerating discovery in genomics and translational research.

Hunter Paxton, South Dakota School of Mines

Dementia is a growing public health concern, with Alzheimer’s disease projected to affect over 13 million Americans by 2050. Early detection of cognitive decline — particularly Mild Cognitive Impairment (MCI), a preliminary stage of Alzheimer’s — is critical for timely intervention. Yet, traditional diagnostic methods (e.g., neuroimaging, CSF analysis) are invasive, costly and often inaccessible.

Recent studies suggest that naturalistic driving behavior may serve as a noninvasive digital biomarker to understand cognitive health. This study investigates whether weekly driving patterns can reliably detect early cognitive decline using a temporally aware convolutional neural network.

We analyze longitudinal driving data from 51 participants, where 10 express early signs of cognitive decline, i.e, MCI. With this study, we show that leveraging temporally aware classifiers can significantly outperform baselines, achieving up to 86% validation accuracy. These findings highlight the potential of AI-driven behavioral monitoring as a scalable, passive tool for preclinical Alzheimer’s detection.

Basanta Adhikari, University of Texas

Learning control objectives from demonstration is essential for enabling robots to acquire complex behaviors from expert guidance. However, many existing learning from demonstration and inverse reinforcement learning (IRL) methods lack structured control interpretations, making it difficult to deploy learned objectives in feedback controllers that require interpretability and performance guarantees. Moreover, in real robotic systems, accurate dynamic models and analytic policy gradients are often unavailable or unreliable. This motivates a framework that can infer control objectives directly from data without explicit system models.

This project develops a model-free inverse reinforcement learning framework for discrete-time linear systems and is validated on leader–follower robotic arms. A model-based approach is first introduced as a theoretical foundation, under the assumption that the expert controller is an optimal control associated with a quadratic cost structure. By exploiting the sensitivity properties of the discrete-time Riccati equation, exact gradients of the feedback gain with respect to the state cost matrix are derived, enabling efficient gradient-based recovery of a cost function whose optimal controller matches the expert policy. Building on the model-based approach, a model-free inverse reinforcement learning approach is developed by learning purely from observed state and control transition data without explicit dynamics models. A data-driven, off-policy inverse Q-learning algorithm is derived, where an equivalent feedback controller is recovered, and the cost parameters are updated using simultaneous perturbation stochastic approximation.

The results demonstrate that inverse reinforcement learning, when formulated in an inverse optimal control (IOC) framework, can effectively recover task objectives from demonstration on robotic manipulators while preserving the interpretability and performance guarantees of control policies.

Matthew Halberg, ����Ӱ��

Large amounts of health care data remain unstructured within clinical notes and documents. Even after text is extracted using optical character recognition (OCR), transforming the contained information into a usable format for clinical and research use is a burdensome task. This study aims to explore the usage of large language models (LLMs) to extract and structure information from provider notes in electronic medical records (EMRs). The cohort is made up of patients with head and neck cancer who underwent radiation and/or chemoradiation therapies. These treatments are often associated with a range of severe and chronic toxicities (mucositis, dysphagia, aspiration, etc.) that can impact patient quality of life and long-term health. In this project, we will use LLMs to identify and classify treatment-related side effects for each patient document, along with their respective severities. Results summarize the types and frequencies of toxicities. This work is the first step toward harnessing and validating the use of LLMs to extract information to support decision-making.