2026 IEEE International Conference on AI and Data Analytics
(ICAD 2026)
June 9 – 10, 2026
Greater Boston, Massachusetts
ICAD 2025 Technical Program
Tuesday, June 24
7:00 AM – 4:00 PM Registration and Check-in
8:00 AM – 8:10 AM Opening
8:10 AM – 8:50 AM Keynote
The Virtues and Vices of Foundation Models
Matthias Scheutz, Karol Family Applied Technology Professor, Tufts University School of Engineering
Matthias Scheutz is the Karol Family Applied Technology Professor of computer science in the Department of Computer Science at Tufts University in the School of Engineering, and Director of the Human-Robot Interaction (HRI) Laboratory and the HRI Masters and PhD programs. Read more >>
8:50 AM – 9:30 AM Keynote
Generative AI: Shaping Tomorrow
Sadid Hasan, AI Lead at Mircrosoft
Sadid Hasan has over two decades of AI R&D experience and currently leads generative AI product development for Office M365 and Azure AIOps AI initiatives as part of Microsoft’s advanced AI Development Acceleration Program. Previously, he was the Executive Director of AI at CVS Health and served as the Senior Scientist and Technical Lead of the AI Group at Philips Research. Read more >>
9:30 AM – 9:50 AM Coffee Break
9:50 AM – 11:50 AM Parallel Tracks
Track 1 Chairs: Reshmi Ghosh (Microsoft) and Amin Ramezani (Harvard Medical School)
Topic: AI for Business Intelligence and Analytics
9:50 AM: Automating Voice of Customer Analysis with AI Workflows Built on GPT-4o-Mini Model presented Kenneth Crowther* (Xylem, INc..); Pietro Aldo Refosco (Xylem, Inc.)
Traditional Voice of Customer (VOC) data analysis methods are labor-intensive and time-consuming, often requiring hundreds of hours for qualitative analysis. This paper introduces an AI-powered workflow based on commercially available large language models that automates the most laborious steps of VOC qualitative data analysis, reducing processing time from hundreds of hours to minutes while maintaining acceptable accuracy and improving consistency. The proposed workflow leverages the generic GPT-4o-Mini model to process and prepare transcript data, identify customer need statements, and organize these statements into a hierarchical structure for further quantitative analysis. The workflow includes stages such as transcript preparation, chunking text for analysis, identifying and scoring needs, and clustering needs into prioritized categories. Evaluation results demonstrate a significant reduction in time and effort, with the AI workflow achieving a 99% reduction in qualitative analysis time compared to traditional methods with sufficient specificity of customer need if the chunk sizes are adequately small. Furthermore, the results help to overcome several typical cognitive and motivational biases of the analysts. The findings highlight the potential of AI to enhance the efficiency and consistency of VOC analysis, providing valuable insights for product development and customer satisfaction improvement.
10:08 AM: Advancing Ethical AI: A Methodological and Empirical Approach to the AI Moral Code presented by Ran Hinrichs* (Norwich University – Vermont)
This paper presents a methodological and empirical framework for the AI Moral Code, based on the Normative, Regulatory, Behavioral, and Conceptual (NRBC) architecture. Analyzing 291 AI ethics documents (2006–2025), it identifies high-frequency values and forecasts emerging trends. The framework translates ethical priorities into system design and governance, offering evidence-based insights and supporting value alignment across sectors such as healthcare, education, justice, and autonomous vehicle technologies.
Topic: Algorithmic Literacy
10:26 AM: Algorithmic Literacy and Digital Privacy in the US: An Exploratory Study Using Data Visualization presented by Haijing Tu* (Indiana State University); Rahul Devajji (Indiana University Bloomington); Tyler Horan (University of Massachusetts Amherst)
This article investigates how algorithmic literacy empowers social media users to control their personal information. Based on secondary data analysis and visualization, this article explores algorithmic bias and digital privacy using the results of the American Trend Panel (ATP) surveys provided by Pew Internet Research. In this article, a set of clear and organized visualizations is used to study digital privacy and the awareness of algorithmic bias. This article also conducts further analyses to explore the correlation between users’ digital privacy scores and their awareness of privacy policies. Based on the findings, participants with higher digital literacy are more likely to intentionally influence or change the content they see on their Facebook news feed. In addition, the better respondents understand the mechanisms of algorithmic-driven content selection, they are more likely to report having control on social media content, whereas respondents with lower score on digital privacy tend to believe they have great control over their personal data collected by the government.
10:44 AM: Parameter-Efficient Adversarial Example Detection and Robustness Enhancement Utilizing Optimized Reverse-Cross Entropy presented by Zirui Fu* (Tufts University); Marco Donato (Tufts University)
Reverse Cross-Entropy (RCE) has shown promise in enhancing adversarial example detection (AED) and robustness capabilities of deep neural network (DNN) models under adversarial attacks. However, high retraining costs and unstable convergence on large-class datasets have limited RCE loss function’s adoption and application. In this work, we refine RCE with temperature scaling and weighted focusing (RCE-TF) to stabilize training on larger tasks. Additionally, we investigate the use of bottleneck residual adapter blocks to efficiently enhance the robustness of pretrained models. Experiments on CIFAR-10 and CIFAR-100 demonstrate that our proposal scales RCE to large dataset classification problems with higher average accuracy compared to conventional loss function. We also achieve superior AED performance and improved robustness up to 10% under multiple adversarial attacks and with specific adapter settings, without prior knowledge to these attacks.
11:02 AM: CB-RML: Dynamic Regret Minimization via Coin-Betting Regularization and Meta-Learning presented by Sourav Dutta* (Ramapo College of New Jersey)
We address the challenge of dynamic regret minimization in non-stationary environments, where traditional online learning algorithms struggle to adapt to shifting data distributions. Existing approaches typically require prior knowledge of environmental dynamics or careful parameter tuning, limiting their practical utility. This paper introduces the \textbf{Coin-Betting Regularized Meta-Learner (CB-RML)}, a novel framework that achieves parameter-free adaptation while maintaining provable regret guarantees. Our key contributions are threefold: (1) a \textit{three-timescale} adaptation mechanism combining per-iteration learning rate updates ($\eta_{t+1}^{(k)} = \eta_0/\sqrt{1 + \sum \|\theta_s^{(k)} – \theta_{s-1}^{(k)}\|^2}$), partition-based restarts, and meta-regularization through $S_t = \sum \|g_t\|^2$; (2) a \textit{parameter-free} design requiring only $\eta_0 > 0$ and $\beta > 0$, eliminating the need for knowledge of $V_T$, $G$, or shift counts; and (3) an automatic \textit{drift detection} mechanism via $\|g_t – g_{t-1}\| > \gamma\sqrt{\log t/|I_k|}$, enabling adaptation to both gradual and abrupt concept shifts. Theoretical analysis demonstrates that CB-RML achieves dynamic regret bounded by $\mathcal{O}(\sqrt{T(1 + V_T)})$, where $V_T$ quantifies environmental variation, while maintaining computational efficiency. Experimental results across synthetic and real-world benchmarks validate our theoretical claims, showing superior performance compared to state-of-the-art baselines in scenarios with unanticipated distribution shifts. The framework’s combination of theoretical rigor and practical applicability advances the frontier of robust online learning in dynamic environments.
Topic: AI in Materials Science & Engineering
11:20 AM: MXene Material Property Prediction via Transfer Learning with Graph Neural Networks presented by Eric Vertina* (WPI); Sajal Chakroborty (WPI); Emily Sutherland (WPI); Nathaniel Deskins (WPI); Oren Mangoubi (WPI)
MXenes are an important class of 2D solid materials with expected novel properties and myriad applications,including efficient energy conversion in batteries and solar cells, environmental and water treatment, supercapacitors, and catalysts. MXenes have the broadest known theoretical chemical space among 2D materials. However, MXenes are challenging to synthesize in the laboratory, and numerical models for simulating MXene properties, such as density functional theory (DFT), remain computationally expensive. We apply transfer learning to train Atomistic Line Graph Neural Network Transfer Learning (ALIGNN and ALIGNNTL)-based models to rapidly predict MX-ene properties, account for crystal graph structure, and leverage additional, non-MXenes data to assist with model initialization. We first train models by incorporating training data from the Computational 2D Materials Database, followed by training these models with MXenes data, to leverage these larger datasets improve model performance. Our best models obtain the following test set mean absolute errors and percent errors, respectively, for the following, individually-predicted MXene properties: Work Function 0.32±0.11 (eV) and 7.5±2.4% error, Magnetic (T/F) of 0.067±0.020 and 6.7±2.0% error, Bulk Modulus 15.5±4.2 (N/m) and 7.3±1.9% error, Shear Modulus 18.5±4.0 (N/m) and 14.6±3.0% error, and Young’s Modulus 38.0±10.7 (N/m) and 12.2±3.3% error. We predict Work Function, Magnetic (T/F), Bulk Modulus, Shear Modulus, and Young’s modulus for each of the 23,857 MXenes in the aNANt database and for each of the 2,784 MXenes presented in Rems et al., a well known paper that claims to lay “the foundation for the materials genomics of MXenes, i.e., MXene genomics.” We use the crystal structure files calculated with DFT from each source as input to the best performing models for each property. Neither of these sources predicted these target properties for their datasets. To our knowledge, no one has predicted the specified target properties for this dataset using the crystal structure information from this dataset as input features such as we predict here. Links to our predicted values for these target properties for all MXenes in these two datasets can be found in the supplemental information.
Track 2 Chairs: Baris Kazar (Oracle) and Nikunj Doshi (Northeastern University)
Topic: Advanced AI Techniques & Emerging Trends
9:50 AM: Lung Cancer Classification using Deep Learning Models for Edge Computing. A Comparative Analysis presented by Sarbagya Shakya* (Eastern New Mexico University); Edgar Ceh-Varela (Eastern New Mexico University); Ivan Sanjaya (Eastern New Mexico University)
Lung cancer remains one of the leading causes of cancer-related deaths worldwide. Early diagnosis and treatment before the cancer spreads can significantly improve survival rates.
Edge devices, including smartphones, wearables, IoT sensors, and embedded systems, are increasingly important for deploying deep learning models in lung cancer detection. In this study, we evaluated three state-of-the-art, compact deep learning models, NasNetMobile, MobileNetV2, and EfficientNetV2, for lung cancer classification using CT scan medical images. Two public datasets, the IQ-OTH/NCCD lung cancer dataset and the Chest CT-Scan images dataset, were employed to assess model performance. EfficientNetV2 emerged as the top performer, achieving accuracies of 98.64% and 89.52% across the datasets, respectively. Its compact size and computational efficiency underscore its potential for deployment in resource-constrained edge computing environments, enabling real-time, scalable diagnostic solutions.
Topic: AI in Finance and FinTech
10:08 AM: ResNet-Enhanced DFSA: A Time-Efficient UHF RFID Inventory System for Large-Scale Applications presented by Heyi Li* (MIT); Sobhi Alfayoumi (UOC); Marta Gatnau Sarret (UOC); Rahul Bhattacharyya (MIT); Joan Melia Segui (UOC); Sanjay Sarma (MIT)
In dense RFID environments, simultaneous responses from multiple tags during reader identification attempts result in signal collisions that compromise successful tag detection. These collisions, predominantly arising from the mismatch between allocated time slots and actual tag population, significantly degrade identification efficiency. The EPC Gen2 protocol employs Dynamic Frame Slotted Aloha (DFSA) to dynamically adjust frame sizes across inventory rounds, yet existing implementations lack precise tag quantity estimation during collision events. This paper presents an AI-enhanced collision resolution framework where a ResNet-based classifier analyzes collided RN16 signals to estimate concurrent responders (up to 10 tags). By integrating this real-time collision cardinality prediction with DFSA’s slot adaptation mechanism, our approach achieves up to 93.3% time saving than conventional implementations through optimized frame size adjustments. The proposed methodology demonstrates particular efficacy in dynamic and dense deployments (>100 tags), establishing a machine learning paradigm for physical-layer signal processing in RFID anti-collision protocols.
Topic: Edge AI
10:26 AM: Machine learning-driven classification of sepsis using influential factor presented by Visalaxi Sankaravadivel* (Numpy Ninja Inc); Kasthuri Indiran (Numpy Ninja Inc); Supratim Das gupta (Numpy Ninja Inc)
The most prominent contemporary technique “Machine learning” performs a significant role in data analysis across various sectors. In the past few decades, these Machine learning techniques have been widely used in the medical industry for analyzing/categorizing various types of diseases. Various machine learning algorithms are used for analysis, especially decision tree and random forest algorithms are one of the regression/classification techniques majorly used for analyzing various types of diseases. Sepsis is one of the life-threatening diseases that occur across various groups of people. Various attributes impact the causes/severity of sepsis. The symptom of sepsis varies based on the age group and previous health conditions. The identified sepsis dataset contains 44 attributes that include the target variable. The most common symptoms include age, rapid heart rate, length of stay in ICU, blood pressure, minerals in the body including lactate, bilirubin, blood union nitrogen, etc., white blood cells, platelets, etc. With the help of the decision tree algorithm and Random Forest, the most influential symptoms were predicted. A decision tree was constructed using entropy and Gini index where ICU length of stay was identified as the first top influential factor, followed by heart rate variation being identified as the next factor from the given dataset. Similarly, the Random Forest classifier also identifies age, and ICU length of stay as the most influencing features from the given dataset. The identified algorithm achieves a higher test accuracy of 98% using both decision tree and random forest.
10:44 AM: Badminton Action Recognition Using Skeleton Data and Optical Flow presented by YuHsuan Tseng (National Tsing Hua University); Kuo-Chin Lin (National Sun Yat-sen University); Che-Rung Lee* (National Tsing Hua University)
The ability to analyze actions in videos is crucial for the automatic understanding of sports. With advancements in action recognition, precise video analysis has become increasingly feasible. Although extensive research has been conducted in the field of action recognition, relatively few studies focus on badminton and broadcast sports videos. In this paper, we propose a two-stream architecture for classifying badminton actions. Our approach leverages two key features—optical flow and skeleton data—as inputs to the two-stream architecture. Given the complexity of badminton movements and techniques, these features are chosen to effectively capture human actions. The extracted features are then processed using a combination of VGG and bi-directional long short-term memory (Bi-LSTM) models for training and classification. Specifically, VGG, a convolutional neural network, is employed in the optical flow stream for feature extraction, while Bi-LSTM is utilized in both streams to capture the dynamic temporal information of video sequences. Our proposed two-stream architecture achieves an accuracy of 94.3\% on our badminton dataset, demonstrating its effectiveness in analyzing broadcast sports videos.
11:02 AM: A Data Pruning Method with Feature Distillation for Improved Computational Efficiency presented by Mike Soricelli (University of Massachusetts Dartmouth); Russell Thompson (NUWC Division Newport); Yuchou Chang* (University of Massachusetts Dartmouth); Christopher Hixenbaugh (NUWC Division Newport)
As the capabilities of Neural Network models grow, so does the cost of power consumption and time, which consequently makes training state of the art models on resource constraint devices more challenging. In this paper, we present a novel method to reduce training costs by performing a version of data pruning, which prunes batches of data based on how far their feature extractions are from a set of feature distilled vectors. We call this new method Data pruning via Feature distillation. We employ a neural network layer, referred to as a feature distilled layer, which maintains a set of vectors which aim to approximate the distribution of extracted features within a neural network. This set of feature distilled vectors dynamically adjusts throughout the training process to accomplish this approximation while also adjusting to changes in the network throughout training. We demonstrate that this method significantly reduces overall training time while maintaining or even improving model performance across various datasets and architectures. Experiments conducted on MNIST, CIFAR10, CIFAR100, and CalTech-256 using ResNet models show performance increases of up to 47.41% speed up and a 40.11% decrease in GPU power consumption, while also maintaining accuracy within 1% of the baseline model. Our work presents as a flexible addition to a neural network which can be added in after feature extraction layers to be used for accelerating training through data pruning.
11:20 AM: Iterative Updating of Digital Twins Using Convolutional Neural Networks: A Framework for Robust Structural Behavior Prediction presented by Zahra Zhiyanpour* (University of Virginia); Zhidong Zhang (University of Virginia); Devin Harris (University of Virginia)
This study introduces a novel iterative updating strategy powered by convolutional neural networks (CNNs) within a digital twin framework for informing structural behavior of infrastructure assets. A two-dimensional (2D) cantilever plate is employed as a benchmark case study for the updating strategy within a digital twin framework. Two distinct surrogate models are the main components of the digital twin framework, which form the basis of this work. The first model is formulated as finite element analysis surrogate, and the training dataset is prepared based on traditional finite element analysis. The second model predicts 2D deformation based on real-world surface images, serving as the ground truth, which is trained on pairs of deformed and reference images of surfaces with speckle patterns. This ground truth deformation data bridges the gap between experimental observations (physical twin) and virtual modeling (digital twin). The core novelty of this study lies in the iterative updating process and the introduction of four specialized CNNs, referred to as “calculators,” designed to refine the inputs for the first surrogate model using the ground truth displacement from the second surrogate model. Each calculator addresses a specific task: 1. Predicting loading in the y direction, given geometry, boundary conditions, loading in the x direction, and ground truth displacement. 2. Predicting loading in the x direction, given geometry, boundary conditions, loading in the y direction, and ground truth displacement. 3. Predicting geometry, given boundary conditions, loading, and ground truth displacement. 4. Predicting boundary conditions, given geometry, loading, and ground truth displacement. These calculators operate simultaneously in an iterative framework, using initial estimates of geometry, boundary conditions, and loading as inputs. Their outputs refine the inputs for subsequent iterations, with the process continuing until convergence. The final refined inputs are used by the first model, and their predictions are compared with the ground truth displacement from the second surrogate model to validate the digital twin. The results demonstrate that this novel approach integrating specialized calculators in an iterative updating process contributes to the robustness and accuracy of digital twins. The proposed updating strategy can pave the way for more reliable structural analysis and predictive modeling through digital twin frameworks, inform efficient infrastructure operation and management, and further contribute to the application and development of digital twin and structural health monitoring techniques.
Track 3 Chairs: Joseph Campbell MIT Lincoln Laboratory) and Soheli Farhana (Harvard University)
Topic: Foundations of AI & Data Analytics
9:50 AM: Multimodal Hateful Meme Detection with Graph Attention Networks and Contextual Cues presented by Dhruv Agarwal* (Northeastern University)*; Hunjun Shin (Northeastern University); Wonhee Lee (Northeastern University); Mahdi Imani (Northeastern University); Naveen Sapavath (Northeastern University)
The classification of hateful memes remains a challenging task due to their multimodal nature, where the interplay of textual and visual elements often conveys implicit or nuanced harmful content. This paper introduces a novel classification framework leveraging Graph Attention Networks (GATs) to model cross-modal relationships between visual and textual components. The proposed method integrates Visual Question Answering (VQA) and image captioning to enhance contextual understanding and refine semantic representations of multimodal data. Each meme is represented as a fully connected graph, where nodes correspond to embeddings derived from visual features, captions, and VQA responses, while GATs dynamically assign importance to these relationships. Experimental evaluation on the HarMeme dataset demonstrates the efficacy of our approach, achieving superior accuracy and AUROC compared to unimodal baselines (ResNet, DistilBERT) and state-of-the-art multimodal models such as Contrastive Language Image Pretraining (CLIP). These results highlight the potential of the proposed GAT-based architectures for improving hateful meme detection and advancing multimodal content analysis. The code is available at https://github.com/Dhruv-2020EE30592/EECE7205-Project-HateMeme.
10:08 AM: Analyzing Deep-Learning Kernel Statistics Through Timm presented by Marika Schubert* (University of Pittsburgh); David Langerman (NSF SHREC); Calvin Gealy (University of Pittsburgh); Evan Gretok (University of Pittsburgh); Alan George (University of Pittsburgh)
The field of deep-learning models for computer-vision tasks has evolved rapidly over the past decade, resulting in a wide variety of models with different computational and memory needs. Before committing to a particular model, engineers must understand the general requirements of models in their domain to choose the proper hardware and algorithms. To identify the constraints for a deep-learning model, this paper proposes deconstructing a repository of models in a given domain down to kernels. These kernels then provide insight into general needs for optimization or acceleration in future research. This method is applied to a repository of deep-learning models for image classification. The models chosen are from the PyTorch-based Timm repository, a widely used library containing hundreds of well-maintained reference implementations of image-classification models. To this end, the Timm Crawler (TiC) was built to aggregate data across this repository. Leveraging this parser, we present a survey of the models contained in this repository relative to kernels used, size of models, and other trends in the field as a whole through the lens of Timm. A key insight from this research is that while the size of the largest models is growing, the median size is not. Within the Timm repository, 92.71 percent of Conv2d layers use 1×1 and 3×3 convolutional kernels. Linear layers are overwhelmingly implemented with biases (91.53 percent) and do not tend to have input/output dimensions that are powers of two (33 percent and 25 percent respectively). These insights into construction indicate that when looking to perform image-classification task, there should be a focus on optimizing or designing accelerators for these types of kernels.
10:26 AM: FIOnA: Feature Invariant Data Augmentation for Small Datasets presented by Winner Kazaka* (Northeastern University); Hangliang Ren (Northeastern University); Tala Talaei Khoei (Northeastern University)
Modern computer vision models face significant challenges when trained on small or domain-specific datasets, where large-scale data collection and labeling are impractical. This paper introduces FIOnA (Feature Invariant Online Aug mentation), a novel data augmentation method grounded in the manifold assumption, which posits that images of the same class form a lower-dimensional manifold within a high-dimensional pixel space. FIOnA generates augmented samples by swapping semantically similar sub-features between images of the same class, and subsequently smoothing to ensure realism. We present a theoretical foundation that demonstrates the label-preserving nature of these feature exchanges. We also detail two feature extraction pipelines Neural Network and Transformer based, as well as two smoothing techniques Diffusion and Generative Adversarial Networks. Experimental evaluation on fine-grained image classification tasks shows FIOnA’s potential to improve accuracy and generalization against baselines by 4.95% and 2.02% above state-of-the-art augmentation techniques, offering an alternative for data-scarce domains and proposing transformative research directions.
10:44 AM: Apriori-Based Antibiotic Association Rule Mining for Optimized Mastitis Treatment in Dairy Cattle presented by Shubhavi Arya (Indiana University, Aria Alessia Research Foundation)*; Minakshi Arya (North Dakota State University, Aria Alessia Research Foundation); Saatvik Arya (University of Washington); Jaibir Singh Arya (4Office of Special Secretary, Animal Husbandry and Fisheries, Government of Haryana, India)
Mastitis is a common and economically significant disease in dairy cattle, necessitating effective antibiotic treatment for optimal management. This study leverages association rule mining and frequent itemset analysis—core techniques in data analytics— to identify the most effective antibiotic combinations for treating mastitis in cows and buffaloes. By analyzing large-scale veterinary data, we extract key associations between antibiotics and assess their efficacy using statistical methods, including t-tests and chi-square tests. The findings provide a computational framework for enhancing evidence-based decision-making in healthcare analytics, demonstrating the potential of AI-driven methodologies in advancing antimicrobial stewardship and optimizing therapeutic outcomes for veterinary practitioners.
11:02 AM: Predicting Cognitive Decline: A Multimodal AI Approach to Dementia Screening from Speech presented by Lei Chi* (The Cooper Union for the Advancement of Science and Art); Arav Sharma (The Cooper Union for the Advancement of Science and Art); Ari Gebhardt (The Cooper Union for the Advancement of Science and Art); Joseph Colonel (Icahn School of Medicine)
Recent progress has been made in detecting early-stage dementia entirely through recordings of patient speech. Multimodal speech analysis methods were applied to the Prediction and Recognition Of Cognitive declinE through Spontaneous Speech (PROCESS) Signal Processing Grand Challenge, which requires participants to use audio recordings of clinical interviews to predict patients as healthy control, mild cognitive impairment (MCI), or dementia and regress the patient’s Mini-Mental State Exam (MMSE) scores. The approach implemented in this work combines acoustic features (eGeMAPS and Prosody) with embeddings from Whisper and RoBERTa models, achieving competitive results in both regression (RMSE: 2.7666) and classification (Macro-F1 score: 0.5774) tasks. Additionally, a novel two-tiered classification architecture is utilized to better differentiate between MCI and dementia. Our approach achieved strong results on the test set, ranking seventh on regression and eleventh on classification out of thirty-seven teams, exceeding the baseline performance.
11:20 AM: Probing a Vision-Language-Action Model for Symbolic States and Integration into a Cognitive Architecture presented by Hong Lu* (Tufts University); Hengxu Li (Tufts University); Prithviraj Shahani (Tufts University); Stephanie Herbers (Tufts University); Matthias Scheutz (Tufts University)
Vision-language-action (VLA) models hold promise as generalist robotics solutions by translating visual and linguistic inputs into robot actions, yet they lack reliability due to their black-box nature and sensitivity to environmental changes. In contrast, cognitive architectures (CA) excel in symbolic reasoning and state monitoring but are constrained by rigid predefined execution. This work bridges these approaches by probing OpenVLA’s hidden layers to uncover symbolic representations of object properties, relations, and action states, enabling integration with a CA for enhanced interpretability and robustness. Through experiments on LIBERO-spatial pick-and-place tasks, we analyze the encoding of symbolic states across different layers of OpenVLA’s Llama backbone. Our probing results show consistently high accuracies (>90%) for both object and action states across most layers, though contrary to our hypotheses, we did not observe the expected pattern of object states being encoded earlier than action states. We demonstrate an integrated DIARC-OpenVLA system that leverages these symbolic representations for real-time state monitoring, laying the foundation for more interpretable and reliable robotic manipulation.
Track 4 Chairs: Soon Wan (Vicor Corporation) and Bruce Shuyue Jia (Boston University)
Topic: AI in Industry & Enterprise Applications
9:50 AM: Optimizing Neural Architectures for Hindi Speech Separation and Enhancement in Noisy Environments presented by Arnav Ramamoorthy* (BITS Pilani)
This paper addresses the challenges of Hindi speech separation and enhancement using advanced neural network architectures, with a focus on edge devices. We propose a refined approach leveraging the DEMUCS model to overcome limitations of traditional methods, achieving substantial improvements in speech clarity and intelligibility. The model is fine-tuned with U-Net and LSTM layers, trained on a dataset of 400,000 Hindi speech clips, augmented with ESC-50 and MS-SNSD for diverse acoustic environments. Evaluation using PESQ and STOI metrics shows superior performance, particularly under extreme noise conditions. To ensure deployment on resource constrained devices like TWS earbuds, we explore quantization techniques to reduce computational requirements. This research highlights the effectiveness of customized AI algorithms for speech processing in Indian contexts and suggests future directions for optimizing edge-based architectures.
Topic: Foundations of AI & Data Analytics
10:08 AM: Introduction to UNET Variant Algorithm based on Transfer Learning presented by Pranshu Tiwari* (Schneider Electrci)
Variety, Sparsity and Vagaries in normality offers tough challenge in anomaly detection. This is further aggravated by varying background on images which creates more difficulty in anomaly detection. To solve this series of labeling is needed to identify object of interest and categorize given the interest and level of detail. This paper presents the FOAK (First of a Kind) metric, incorporating the Intersection over Union (IoU) method, and integrating a unique autoencoder connector for transfer learning with the U-Net algorithm. The motivation for this innovation stems from the need for efficient and objective anomaly detection in visual inspections, particularly in manufacturing settings, to address the limitations of manual checks. The proposed solution aims to provide reliable and scalable anomaly detection, leveraging semantic segmentation for precise object identification. The approach involves a detailed comparison of semantic segmentation methods, leading to the development of a new semantic segmentation technique. The invention also includes a novel double autoencoder layer within the U-Net algorithm, enhancing its ability to process and classify images. The novel components of the invention are evaluated against existing techniques, demonstrating superior performance in object identification using categorical cross entropy. The results highlight the potential of the proposed approach in addressing semantic segmentation tasks, particularly in scenarios with a limited number of distinct instances within images.
Topic: AI for Business Intelligence and Analytics
10:26 AM: Bridging Augmented Reality and AI for Secure and Personalized Educational Experiences presented by Seshagirirao Lekkala* (Cisco systems Inc); Priyanka Gurijala (Microsoft Corporation); Sambu Patach Arrojula (Samsung Research America); Deepak Bhaskaran (Cisco Systems Inc)
The integration of Augmented Reality (AR) and Artificial Intelligence (AI) in education has emerged as a transformative approach to enhancing student engagement and personalized learning experiences. AI has accelerated the growth of AR dramatically. This paper explores the synergistic potential of AR and AI technologies, highlighting their capacity to create an education environment that cater to diverse requirements of education. It is found that AR applications significantly boost student motivation and knowledge retention, while AI-driven adaptive learning systems boost up academic performance by tailoring educational content to specific learners. However, there are some challenges related to data privacy, security, and equitable access to technology remain critical concerns that must be addressed to fully realize the benefits of these innovations. By examining existing literature and presenting a comprehensive framework for implementation, this study underscores the importance of ethical considerations and the evolving role of educators in AI-augmented environments. The findings emphasize the importance of ongoing research to develop effective strategies that bridge technology and pedagogy, resulting in secure, personalized, and engaging educational experiences.
Topic: Monitoring and Evaluation
10:44 AM: A Delphi-Driven Ontology for Integrating Big Data in Monitoring and Evaluation presented by Tinashe Madamombe* (Durban University of Technology)
The integration of M&E systems presents transformative opportunities for enhancing evidence-based decision-making. However, the absence of a standardized framework for managing and interpreting Big Data in M&E contexts hampers its full potential. This study addresses this critical gap by developing and validating a Big Data ontology tailored for M&E applications. Utilizing the Delphi method, the research engaged a panel of 14 experts from diverse fields, including M&E, Big Data analytics, and ontology development, to iteratively refine and validate the framework. The study’s methodology follows a three-phase process: (1) constructing an initial ontology through an extensive literature review and conceptual analysis, (2) validating and refining the framework through multiple rounds of expert consensus, and (3) finalizing the ontology based on systematic feedback integration. Key components of the ontology include core classes such as Projects, Indicators, Measurements, Outcomes, and Big Data, structured to enhance data interoperability, scalability, and analytical precision within M&E practices. Validation results, measured using Kendall’s coefficient of concordance (W=0.82), indicate strong expert consensus on the ontology’s theoretical soundness, practical utility, and technical accuracy. This study makes significant contributions by providing a structured framework that standardizes Big Data integration in M&E, facilitating improved decision-making, transparency, and scalability. The findings offer valuable insights for policymakers, researchers, and practitioners, with implications for cross-sectorial applications and digital transformation in evaluation practices. Future research should explore empirical validation through case studies and adaptation of the ontology for broader domains beyond M&E.
Topic: AI in Industry & Enterprise Applications
11:02 AM: Balancing Automation and Human Oversight in Healthcare AI presented by Swagata Ashwani* (Boomi); Meetu Malhotra (Harrisburg University of Science and Technology); Shriya Agarwal (University of the Cumberlands); Brinda Gurusamy (University of California, Berkeley); Divya Karthikeyan (Montclair State University)
AI has influenced the healthcare domain significantly, improving diagnostic precision and predictive modeling. However, to ensure safety, trust, and reliability in AI-driven decision-making, there needs to be a balance between automation and human oversight. This review will discuss how simulated human feedback can be utilized to improve model performance by identifying and correcting low-confidence predictions. We train baseline models, Logistic Regression and ResNet-50, on MIMIC-IV and CheXpert datasets, respectively, while introducing mechanisms of oversight for refining the predictions. This reflects in various evaluation metrics used here, like AUC-ROC, precision, and recall. Our results underline the necessity of adaptive control mechanisms in AI for healthcare applications.
11:20 AM: Evaluating Accuracy in Large Language Models: Benchmarking Corrective RAG vs. Naive Retrieval Augmented Generation Approach presented by Rajendra Gangavarapu* (GSU); Venkata Moparthi (university of illinois); Aswath Ram Adayapalam Srinivasan (IIT)
Retrieval-Augmented Generation (RAG) has emerged as a promising approach to mitigate the limitations of Large Language Models (LLMs) in generating factually accurate and consistent text. The main focus of this technical survey is on correct answers as the key performance indicator (KPI) for comparing two well-known RAG methods: Naive RAG and Corrective retrieval augmented Generation (CRAG). Naïve RAG exhibits a strong dependence on the relevance of retrieved documents, resulting in suboptimal performance when retrieval quality is compromised. CRAG, on the other hand, adds new features to improve robustness and adaptability, such as a retrieval evaluator, large-scale web searches, and a decompose then recompose algorithm. We introduce the Comprehensive RAG Benchmark (CRAG), which encompasses a diverse set of question-answer pairs spanning multiple domains, categories, entity popularities, and temporal dynamics, to facilitate a comprehensive evaluation of RAG models’ performance in generating correct answers. Experiments show that adding RAG to LLMs makes a big difference in how many correct answers you get, with CRAG consistently beating Naive RAG. Nevertheless, RAG models, including CRAG, demonstrate lower answer correctness when confronted with questions pertaining to highly dynamic, less popular, or more complex facts. These results make it clear that more research and development is needed to make RAG models more reliable and able to give correct answers in a wider range of situations.
11:40 PM – 1:30 PM Lunch and Concurrent Workshops
11:50 AM – 12:40 PM Session 1 Concurrent Workshops:
- TraML Meets AI in Space
- Transformer Neural Networks: A Pedagogical Overview
12:40 PM – 1:30 PM Session 2 Concurrent Workshops:
- Beyond Accuracy: Exploring Responsible AI in Emotion Recognition and Mental Health
- Interpretable Model Learning for Taskable AI Systems
1:30 PM – 2:30 PM Panel Discussion
Responsible Generative AI: Ensuring Successful Real World AI Deployments
Moderator: Dr. Mark Maybury, VP Lockheed Martin, Moderator
Dr. Mark Maybury is the vice president, Commercialization, Engineering & Technology for Lockheed Martin, responsible for leading efforts to commercialize dual-use products and services across the corporation. Dr. Maybury’s prior roles include first chief technology officer for Stanley Black & Decker, Chief Scientist of the U.S. Air Force, Chief Technology and Chief Security Officer at MITRE and Director of the National Cybersecurity FFRDC. Read more >>
2:30 PM – 3:50 PM Parallel Tracks
Please click the Track box to view the presentations that will be given for that track.
Track 1 Chairs: Reshmi Ghosh (Microsoft) and Amin Ramezani (Harvard Medical School)
Topic: Responsible AI & Ethics
2:30 PM: Leveraging Large Language Models for Requirements Generation: An Evaluation through Systems Engineering Guidelines presented by Joel Stein (MITRE); Tomi Esho (MITRE); Jyotirmay Gadewadikar* (MITRE)
Identifying business needs and system requirements is a crucial task in the systems engineering life cycle. With recent advancements in large language models (LLMs), there is growing interest in their potential to enhance the requirements engineering process. This study explores the capabilities of LLMs in generating and clustering requirements by developing a dedicated test platform. Our findings indicate that LLM-based tools can significantly aid in requirements engineering, achieving strong results in coverage and cluster relevance. Ongoing efforts focus on prototyping new applications and expanding the evaluation framework to further assess the impact of LLMs in this domain.
2:48 PM: Shared Control with Black Box Agents using Oracle Queries presented by Inbal Avraham (Bar Ilan University); Reuth Mirsky* (Tufts University)
Shared control problems deal with an agent learning to act in collaboration with other agents or systems. When learning a shared control policy, often a short communication between the agents can significantly reduce running times and improve the system’s accuracy.
We extend the shared control problem to include the ability to directly query a cooperating agent. We consider two types of potential responses to a query, namely oracles: one that assumes that the oracle has a perfect knowledge of the shared system and can provide the learner with the best action it should take, even when that action might be myopically wrong (a teacher), and one with a bounded knowledge limited to its own part of the system (an expert).
Given this additional information channel, this work presents several heuristics for choosing when to query: reinforcement learning-based, utility-based, and entropy-based. These heuristics are aimed at reducing the overall learning cost of a system. Empirical results on two domains show the benefits of querying to learn a better control policy and demonstrate the tradeoffs between the proposed heuristics.
3:06 PM: Identification of Factors Correlating to Patient Appointment No-shows Using Deep & Machine Learning presented by Khald Aboalayon* (Clark University); Nazrinbanu Nagori (Clark University); Kunal Malhan (Clark University); Emre Tokgoz (Clark University); Hassan Musafer (Clark University); Corey Kiassat (NC State University)
Patient no-shows, scheduled but unattended medical appointments, pose a significant challenge in healthcare, impacting patient healthcare by disrupting the necessary treatment. Medical appointment no-shows disrupt healthcare providers’ schedules that leads to suboptimal patient care. Such issues also lead to increased costs on health care, particularly in the medical field where resources are costly and in great demand. This study explores the factors influencing patient no-shows experienced by a multi-specialty clinic in the USA through the analysis of a dataset with more than 38000 medical records collected over a two-month period in 2019. Using descriptive modeling techniques, we examine key attributes such as patient demographics, appointment types, and scheduling patterns to identify trends associated with appointments missed by the patients. Our analysis reveals 22% of the appointments missed, with significant variations across different ethnic groups, appointment types, and days of the week. Notably, the Hispanic or Latino patient cluster exhibited a higher no-show rate in comparison to the others. Predictive modeling techniques used included Random Forest, Support Vector Machine (SVM), ANN, Gradient Boosting (GBM), and XGBoost to assess the feasibility of forecasting no-shows. The findings provided in this work provide an insight into appointment trends and disparities that may impact decision makers’ approach to patient appointment scheduling.
Track 2 Chairs: Baris Kazar (Oracle) and Nikunj Doshi (Northeastern University)
Topic: Data Analytics
2:30 PM: Trends in US Healthcare Data Breaches presented by Li Xu* (The University of Arizona)
Healthcare data breaches remain the most costly among industries, threatening patient privacy, organizational integrity, and public trust. This study analyzed healthcare data breaches reported to the United States (US) Department of Health and Human Services (HHS) Office for Civil Rights (OCR) between 2019 and 2024 to uncover trends, geographic distributions, breach types and locations, and response strategies. Findings indicate an increasing frequency and severity of breaches, with large-scale incidents disproportionately affecting vast populations. Geographic and temporal patterns reveal that populous and affluent states—such as California, Texas, and New York—consistently face higher risks, while other regions experience intermittent surges. The study also highlights a shift in cybercriminal tactics, with network servers and electronic records emerging as primary targets, signaling broader technological trends and a decline in physical breaches. Despite advances in artificial intelligence and machine learning for cybersecurity, human error remains a critical vulnerability. These findings underscore the need for integrated solutions combining regulatory enforcement, technological innovation, and human-centered strategies. This study provides actionable insights to mitigate financial and reputation damages, fostering improved cybersecurity resilience in the healthcare sector.
Topic: Health Care
2:48 PM: A Path to Improved Fetal Cardiovascular Health Outcomes Using Machine Learning presented by Christina Quin* (Lincoln-Sudbury Regional High School)
Reduction of child mortality is a significant focus of the UN’s Sustainable Development Goals (SDGs). The UN seeks to eliminate preventable neonatal and under-5 deaths by 2030 [1]. Maternal mortality, which accounted for 295,000 deaths during and following pregnancy and childbirth in 2017, is closely linked to child mortality. Most of these deaths 94 percent occurred in low-resource settings and were preventable [2]. Cardiotocograms (CTG) offer a simple and cost-effective method to assess fetal health, which is crucial in preventing child and maternal mortality. CTG equipment works by sending ultrasound pulses and analyzing the responses to monitor fetal heart rate (FHR), fetal movements, uterine contractions, and other parameters in the dataset. Using CTG data, machine learning models—Multinomial Logistic Regression, Random Forest, Gradient Boosting and Fully Connected Feedforward Neural Network (FCNN) —were employed to classify fetal health into three categories: Normal, Suspect, and Pathological. Strategic feature preprocessing and class-balancing techniques were implemented to enhance model performance. Results indicate that FCNN achieved the highest overall accuracy, while Gradient Boosting provided superior recall for pathological cases. Additionally, FCNN demonstrated strong recall for the suspect category, ensuring accurate identification of at-risk cases. These findings suggest that FCNN is an effective model for fetal health classification, with Gradient Boosting serving as a strong alternative for prioritizing recall of clinical scenarios. By minimizing observer dependency, the study aims to mitigate unnecessary medical interventions while providing a consistent, accurate, and cost-effective approach to neonatal health assessment [3].
3:06 PM: Natural Language Interface for Queries on Databases with Sensitive Information presented by Suli Adeniye* (Arizona State University); Faisal Al-Atawi (Arizona State University); Arunabha Sen (Arizona State University)
The analysis of large datasets poses significant challenges for users unfamiliar with structured query languages such as SQL or Cypher. These users, often from fields like social sciences or law enforcement, require intuitive and secure access to complex databases. Additionally, datasets in domains such as criminal justice, healthcare, and education often contain sensitive information, necessitating robust privacy protections. To address these challenges, we propose a privacy-preserving natural language interface for querying graph databases. Our framework integrates optical character recognition (OCR) techniques to extract structured data from scanned and unstructured documents, storing it in a Neo4j graph database to efficiently model entity relationships. To facilitate user interaction, we leverage a Large Language Model (LLM)-powered query translation module that converts natural language queries into Cypher, eliminating the need for technical expertise.
To enhance privacy, our framework pseudonymizes sensitive entities using entity masking before query translation, ensuring that personally identifiable information within the user queries is never exposed to the LLM. Additionally, the Neo4j database is entirely separated from the LLM, meaning that the LLM only has access to the database schema and not any stored data. This ensures that query processing occurs in a controlled, authenticated environment, mitigating risks of data leakage. We also enforce schema-aware query translation, which significantly improves accuracy by guiding the LLM with structured database constraints.
We evaluate our system using real-world criminal records and the publicly available Text2Cypher dataset, demonstrating high logical and execution accuracy. Computational performance analysis shows that our privacy mechanisms introduce only minimal latency overhead (∼5-8%), making them suitable for real-time applications. Our results confirm that this framework effectively balances accessibility, accuracy, and privacy, making it applicable to sensitive domains where secure, user-friendly data retrieval is essential.
Track 3 Chairs: Joseph Campbell MIT Lincoln Laboratory) and Soheli Farhana (Harvard University)
Topic: Foundations of AI & Data Analytics
2:30 PM: Predicting Poverty in the US Using Machine Learning on Demographic and Socioeconomic Data presented by Qiaorui Zhang (Newton North High School); David Nizovsky (Vanderbilt University); Tingying Helen Zeng (Innobridge Institute); Mikhail Shalaginov* (MIT)
Poverty is a pervasive issue worldwide, including in the US, where millions face financial hardship and limited access to essential resources. Poverty prediction is crucial for developing interventions that alleviate hardship and reduce cyclical inequality. We used supervised machine learning models to predict poverty in the US, offering a more efficient and accessible approach than traditional methods. Four binary-classification models were trained on data from the 2022 American Community Survey Public Use Microdata Sample. The Extreme Gradient Boosting (XGBoost) algorithm performed best, yielding an accuracy of 88.80% and a F1-score of 88.37%. Results identify employment, education, mobility status, relationship to householder, and age as the most indicative factors of poverty. This study demonstrates that classification algorithms are valuable tools for identifying and understanding poverty across America.
2:48 PM: Lockheed Martin AI Factory: Generative AI and MLOps for Engineering, Enterprise and Edge presented by Mark Maybury* (Lockheed Martin)
This article reports on the rapid creation and deployment at scale of hundreds of Large Language Model (LLM) applications, highlighting several across enterprise, engineering and edge use cases. This outcome was accelerated by the creation of an open architecture, secure and scalable generative AI Factory, a platform that empowers thousands of developers and over 50,000 end users across a diverse set of data types and use cases throughout our global enterprise. The approach reveals how to affordably deploy generative AI to create value securely at scale.
Topic: Monitoring and Evaluation
3:06 PM: Mitigating Biased, Brittle and Baroque Generative AI presented by Mark Maybury* (Lockheed Martin)
The rapid adoption of Generative AI across multiple use cases has introduced new risks and attack surfaces across a range of important applications. While the increasing competency of Large Language Models (LLMs) promises high value, responsible AI requires the mitigation of potentially biased, brittle, and baroque operations which undermine trust and carry the potential of causing harm. Moreover, active adversaries will exploit LLM weaknesses to undermine confidentiality, integrity, and availability. Encompassing a survey of current LLM performance and risk literature, this paper presents a framework for understanding the current state of the art and a complementary path forward. In particular, the framework enables designers, users, and policymakers to understand and assess categorized risks and suggests methods to mitigate weaknesses during design and deployment in order to create LLMs that are less likely to fail and more resilient to attacks. It also suggests methods to use existing but imperfect LLMs in a responsible manner.
3:42 PM: Gap the (Theory of) Mind: Sharing Beliefs About Teammates’ Goals Boosts Collaboration Perception, Not Performance presented by Yotam Amitai* (Technion); Reuth Mirsky (Tufts University); Ofra Amir (Technion)
In human-agent teams, openly sharing goals is often assumed to enhance planning, collaboration, and effectiveness. However, direct communication of these goals is not always feasible, requiring teammates to infer their partner’s intentions through actions. Building on this, we investigate whether an AI agent’s ability to share its inferred understanding of a human teammate’s goals can improve task performance and perceived collaboration. Through an experiment comparing three conditions—no recognition (NR), viable goals (VG), and viable goals on-demand (VGod)—we find that while goal-sharing information did not yield significant improvements in task performance or overall satisfaction scores, thematic analysis suggests that it supported strategic adaptations and subjective perceptions of collaboration. Cognitive load assessments revealed no additional burden across conditions, highlighting the challenge of balancing informativeness and simplicity in human-agent interactions.
These findings highlight the nuanced trade-off of goal-sharing: while it fosters trust and enhances perceived collaboration, it can occasionally hinder objective performance gains.
Track 4 Chairs: Soon Wan (Vicor Corporation) and Bruce Shuyue Jia (Boston University)
Topic: AI in Industry & Enterprise Applications
2:30 PM: AI-Driven Architectures for Real-Time Decision-Making in Autonomous Vehicles presented by Naveen Naik Sapavath* (Northeastern University); Sudheer Amgothu (Pega System); Suraj Patel Muthe Gowda (Northeastern University)
In this paper, we design, develop, and evaluate an AI-based architecture to enhance the intelligence and decision-making capabilities of autonomous vehicles. Autonomous vehicles represent a critical emerging application that relies on network services with stringent requirements such as low latency and high data throughput to enable intelligent and timely decisions. The proposed architecture integrates four key components: 1) AI-based Sensors for gathering and interpreting environmental data; 2) Inference Models and Risk Assessment for predicting potential hazards and making proactive decisions; 3) Navigation and Decision-making systems for optimal route planning and control; and 4) Real-Time Data Processing and Event-Driven mechanisms for immediate response to dynamic driving conditions. These components are developed as microservices, allowing for scalable, flexible, and efficient operation. The dynamic interactions between these services ensure the autonomous vehicle can respond to real-world conditions in real-time, leading to safer, more efficient driving experiences. Our evaluation demonstrates the effectiveness of this architecture in managing complex driving scenarios while maintaining system performance and reliability
2:48 PM: AI-Driven Prescriptive Analytics for Hydrate Mitigation in Offshore Petroleum Production presented by Mateus Fernandes* (Petrobras); Rafael Rabelo (Petrobras); Eduardo Gildin (Texas A&M University); Marcio Sampaio (University of São Paulo)
In offshore petroleum systems, one of the main challenges in production management and flow assurance is preventing blockages caused by natural gas hydrates. These crystalline structures form when water and natural gas combine under specific temperature and pressure conditions, leading to well unavailability and requiring significant resources for restoration. In this study, we evaluate hydrate mitigation strategies and their associated economic aspects using an AI-driven prescriptive analytics approach. We propose a framework based on the probabilistic modeling of influencing events, such as unscheduled shutdowns and the probability of hydrate formation based on the duration of fluid stagnation in subsea pipelines. This probabilistic analysis is combined with deterministic data, including production forecast curves, the duration and priority order of mitigation actions, and economic data related to revenues and expenditures. Using this data, we deploy Monte Carlo simulations alongside a rule-based system to assess differences in cumulative production, blockage probabilities, and net present value (NPV) under various scenarios involving production strategies and improvements in hydrate prevention systems. We demonstrate the practical applicability of this framework through three real-world case studies involving an offshore production platform operating in the Brazilian pre-salt, where the results provide valuable insights for decision-making aimed at maximizing value.
3:06 PM: Physics-Informed Deep Learning Prediction of Completion Offsets for Automated Cased-hole Petrophysical Analysis presented by Saad Omar* (Schlumberger-Doll Research); Wail Benrabh (Schlumberger-Doll Research); Jeffery Miles (Schlumberger-Doll Research); Laurent Mosse (Schlumberger-Doll Research)
A deep learning data processing pipeline is developed which automates spectral yield adjustments in neutron induced gamma-ray spectroscopy for cased-hole formation evaluation logging in Oil and Gas wells. Traditional workflows depend on a manual-intensive iterative exercise led by subject matter experts to find cased-hole completion (casing and cement) offsets thereby isolating formation-specific yields. Using a novel mixed-data transfer learning technique, models are trained on combined laboratory and field datasets, incorporating spectral yields, formation properties, geometric parameters (e.g. cement/casing areas), and engineered feature combinations. Using the cement fraction method, calcium contributions from cement were systematically added to the laboratory dataset, extending its applicability to carbonate formations. Conventional transfer learning from laboratory data trained model to field data proved ineffective due to catastrophic forgetting; instead, training on a mixed dataset from the outset achieved robust performance over diverse controlled and field datasets by consistently achieving desired accuracies of 0.01 for cement capture (silicon and calcium) offsets, 0.03 for casing capture (iron) offset and normalized value of 10 for inelastic calcium offset. This work advances automated formation evaluation front by embedding first-principles and domain knowledge with scalable deep learning solutions, enhancing efficiency and reproducibility in subsurface compositional analysis.
3:24 PM: High Accuracy Preserving Regression-Based Physics Inversion Workflow Deployment on 8-bit Integer Computing Hardware presented by Saad Omar* (Schlumberger-Doll Research); Ossama Chrifi (Schlumberger-Doll Research); Mehdi Hizem (Houston Reservoir Lab)
Deep neural networks (DNNs) enable real-time regression-driven inversion in edge AI systems across various domains, including geophysical exploration. However, quantizing such models to low-bit compute precision for fast and power-efficient embedded deployment is challenging, especially in continuous-valued regression tasks that demand high prediction accuracy. In this work, we deploy a multitask learning (MTL) DNN model on Neural Processing Unit (NPU) for predicting geophysical properties (e.g formation density, photoelectric factor, and mud properties). By combining customized per layer quantization-aware training (QAT) and noise-augmented synthetic data, we achieve near-floating-point accuracy while drastically reducing inference latency. Experimental results show over 17× speedup on the NPU compared to existing inversion algorithm.
3:50 PM – 4:10 PM Coffee Break
4:10 PM – 5:10 PM Parallel Tracks
Please click the Track box to view the presentations that will be given for that track.
Track 1 Chairs: Reshmi Ghosh (Microsoft) and Amin Ramezani (Harvard Medical School)
Topic: Generative AI and Creative Applications
4:10 PM: Gradient Boosting Decision Trees on Medical Diagnosis over Tabular Data presented by Ayberk Yarkin Yildiz* (Northeastern University); Asli Kalayci (Worcester Polytechnic Institute)
Medical diagnosis is a crucial task in the medical field, in terms of providing accurate classification and respective treatments. Having near-precise decisions based on correct diagnosis can affect a patient’s life itself, and may extremely result in a catastrophe if not classified correctly. Several traditional machine learning (ML), such as support vector machines (SVMs) and logistic regression, and state-of-the-art tabular deep learning (DL) methods, including TabNet and TabTransformer, have been proposed and used over tabular medical datasets. Additionally, due to the superior performances, lower computational costs, and easier optimization over different tasks, ensemble methods have been used in the field more recently. They offer a powerful alternative in terms of providing successful medical decision-making processes in several diagnosis tasks. In this study, we investigated the benefits of ensemble methods, especially the Gradient Boosting Decision Tree (GBDT) algorithms in medical classification tasks over tabular data, focusing on XGBoost, CatBoost, and LightGBM. The experiments demonstrate that GBDT methods outperform traditional ML and deep neural network architectures and have the highest average rank over several benchmark tabular medical diagnosis datasets. Furthermore, they require much less computational power compared to DL models, creating the optimal methodology in terms of high performance and lower complexity.
4:28 PM: Evaluating the Feasibility of Running AI Large Language Models Locally: Performance, Cost, and Strategic Insights presented by Adrian Besimi* (South East European University)
This article evaluates the use of three different LLMs in local deployment scenarios and during experiment phase it measures time of execution, memory consumption and power consumption that results in cost measurements. The article starts by providing backgroun information from the literature, moving into experimental stage and then providing the results. The article discusses the findings and provides feedback in a form of strategic decission that SMEs and or individual researchers can apply.
Track 2 Chairs: Baris Kazar (Oracle) and Nikunj Doshi (Northeastern University)
Topic: Applied Machine Learning
4:10 PM: Energy Forecasting in High Performance Computing Datacenters Using Machine Learning presented by Leslie Horace (Georgia Institute of Technology); Christopher Stokes (Coastal Carolina University); Craig Walker (Coastal Carolina University); Anvitha Ramachandran (University of Southern California); William Jones* (Coastal Carolina University); Nathan DeBardeleben (Los Alamos National Laboratory); Steven Senator (Los Alamos National Laboratory)
The size and scope of large-scale High-Performance Computing (HPC) computer systems have continued to grow, especially with respect to power density and overall capacity as increasingly performant high-end General-Purpose Graphics Processing Units (GPGPUs) have paved the way for large-scale deep learning models across a range of application domains. Data centers and their operators must contend with increasingly constrained power caps at the same time as effectively managing overall energy consumption. Often, large data centers need to provide estimates of expected future demand and energy across a variety of timescales, including monthly, weekly, daily, and in some cases, even hourly forecasting. There can be significant financial incentives from the local power grid to accurately predict these energy needs. These incentives are often in the form of lower rates and reduced penalties for either over- and under-consumption. In this work, we make use of several strategies to make these predictions, ranging from simple bulk statistical models to more sophisticated machine learning (ML) techniques that leverage historical energy usage, past and present job characteristics, as well as dynamic system queue pressure to improve prediction accuracy.
In this paper, we discuss these techniques, evaluate their costs, dependencies, assumptions and overall effectiveness using actual workload and energy data from a cluster of interest to the United States Department of Energy. We demonstrate that relatively accurate real-time datacenter energy predictions can be obtained by making use of a modest set of input training data and that these predictions can be refined to be even more accurate given additional information provided by a conservative backfilling scheduler.
4:28 PM: Evaluating Convolutional Neural Networks for Synthetic Image Detection in the Frequency Domain presented by Sami Nourji (Brown University); Tanay Subramanian (Brown University); Sujith Pakala (Brown University); Everest Yang* (Brown University)
This paper evaluates the addition of Fourier Transforms into CNN models designed to distinguish real images from AI-generated ones, addressing challenges posed by the rise of hyper realistic content produced by AI. Unlike prior studies that apply frequency-domain analysis to general image classification tasks, our research specifically focuses on AI-generated image detection using the CIFAKE dataset. We implement a CNN architecture with Fourier Transform features to classify synthetic images. We hypothesize that combining frequency and spatial domain information improves CNN-based detection of AI-generated images. We ultimately disprove this through two experiments that show that our best-performing baseline CNN achieved a testing accuracy of 98.58%, while our Fourier-based model reached an accuracy of 98.50%. Our findings highlight that incorporating Fourier features into the detection pipeline provides valuable insights, although the overall accuracy depends mostly on the CNN architecture. This research contributes to the growing field of digital authenticity by critically assessing the role of Fourier-based enhancements, demonstrating that spatial and frequency-domain integration does not necessarily yield superior results in detecting AI-generated images.
4:46 PM: Advanced Preprocessing Techniques for Transaction Data Analysis presented by Eirini Lagiou (University of Ioannina); Anastasia Trantza (Athena Research Center); Jeries Besharat (University of Ioannina); Voula Georgopoulos* (University of Patras); Chrysostomos Stylios (University of Ioannina)
The analysis of transaction data is often impeded by challenges such as noise, inconsistencies, high dimensionality, and missing values, which obscure valuable insights. Existing preprocessing techniques, while effective in isolated scenarios, often struggle to address the complexity and heterogeneity of large-scale transaction datasets, particularly when dealing with dynamic, high-frequency, and multi-source data. This paper presents a comprehensive preprocessing framework that integrates advanced techniques—such as data cleaning, transformation, feature engineering, and dimensionality reduction—to enhance data quality, consistency, and usability. Methods including Principal Component Analysis (PCA), Singular Value Decomposition (SVD), and clustering algorithms are employed to address data quality issues and improve pattern recognition. Unlike previous approaches, the proposed framework systematically addresses both data inconsistency and scalability, demonstrating superior performance in enhancing clustering accuracy, identifying transaction trends, and revealing user behavior patterns. Experimental results confirm that our approach significantly improves data structure clarity, supporting more robust decision-making. These findings highlight the necessity of advanced preprocessing in transaction data analysis and lay the groundwork for future AI-driven methodologies aimed at improving data handling and analytical performance.
Track 3 Chairs: Joseph Campbell MIT Lincoln Laboratory) and Soheli Farhana (Harvard University)
Topic: Quantum Machine Learning
4:10 PM: Quantum Diffusion Models for Few-Shot Learning presented by Ruhan Wang* (Indiana University); Ye Wang (Mitsubishi Electric Research Laboratories); Jing Liu (Mitsubishi Electric Research Laboratories); Toshiaki Koike-Akino (Mitsubishi Electric Research Laboratories)
Modern quantum machine learning (QML) methods involve the variational optimization of parameterized quantum circuits on training datasets, followed by predictions on testing datasets. Most state-of-the-art QML algorithms currently lack practical advantages due to their limited learning capabilities, especially in few-shot learning tasks. In this work, we propose three new frameworks employing quantum diffusion model (QDM) as a solution for the few-shot learning: label-guided generation inference (LGGI); label-guided denoising inference (LGDI); and label-guided noise addition inference (LGNAI). Experimental results demonstrate that our proposed algorithms significantly outperform existing methods.
4:28 PM: A Hybrid Pruning-Quantization Framework for Compact and Efficient Spiking Neural Networks presented by Alissa Kane (University of Massachusetts Dartmouth); Felipe Marcelino (NUWC Division Newport); Anton Spirkin (NUWC Division Newport); Yuchou Chang* (University of Massachusetts Dartmouth)
Spiking Neural Networks, or SNNs, are event driven and suitable for energy-efficient and high-throughput computing. Pruning unnecessary synaptic connections or neurons helps reduce model complexity, decreasing computation and memory requirements whilst preserving inference accuracy. Quantization is also effective in reducing model size by mapping high-precision weights to lower bit-width representations. In this paper, we propose three methods of reducing model size and complexity including pruning and quantization, as well as a hybrid pruning and quantization method. We aim to use these methods to significantly reduce the SNN model size while still maintaining high predictive performance. Our experimental results on oceanographic data indicate that the methods achieve competitive accuracy with a substantial decrease in model size.
4:46 PM: The Effects of Noise on Multimodal Spiking Neural Networks presented by Jacob Fronzaglia (University of Massachusetts Dartmouth); Anton Spirkin (NUWC Division Newport); Felipe Marcelino (NUWC Division Newport); Yuchou Chang* (University of Massachusetts Dartmouth)
Spiking Neural Networks are a relatively new type of energy-efficient artificial intelligence model. Studies on this model include unimodal models and multimodal models. Existing research studies multimodal spiking neural networks (SNNs) which fuse different types of data such as images and audio into one model. However, the effects of noise on the multimodal model are not thoroughly investigated. Noise types and noise levels in each data modality may influence multimodal SNN performance. In this paper, the proposed method is a new framework to study the effects of noise for multimodal SNNs on the classification task. Preprocessing techniques, the insertion of audio and image noise, and the Leaky-Integrate-and-Fire neurons are explored. Experimental results show that the multimodal SNN outperforms its unimodal counterparts. Moreover, some types of audio and images performed better than others, as well as some noise levels performing better than others. The datasets were all frames of images and snippets of audio extracted from videos. The simulated noise was generated from software, which shows that this method can be improved with real noise from real-world data acquisition in future work.
Track 4 Chairs: Soon Wan (Vicor Corporation) and Bruce Shuyue Jia (Boston University)
4:10 PM: Comparative Analysis of Bitcoin Price Movement Prediction Using ARIMA and FBProphet presented by Veronica Dwiyanti Witak Keluli* (Bina Nusantara University); Tuga Mauritsius (Bina Nusantara University)
The rise of Bitcoin and other cryptocurrencies has transformed the financial landscape, especially emerging markets in Indonesia, where adoption rates have grown significantly in recent years. With Indonesia ranked among the top countries in terms of global crypto usage, the demand for innovative strategies to predict Bitcoin price movements has increased. This study compares the ARIMA and Facebook Prophet models for forecasting Bitcoin price trends using historical data from Indodax, one of Indonesia’s largest cryptocurrency exchanges, covering the period from 2019 to 2024. Pre-processing involved handling missing data and applying feature engineering techniques such as moving averages and rolling statistics. Results indicate that ARIMA outperformed FBProphet in accuracy, achieving an RMSE of 26,896,558 and MAPE of 1.81%. While FBProphet excelled in capturing seasonal patterns despite its higher RMSE and MAPE, ARIMA demonstrated superior precision but struggled with high market volatility. This study highlights the complementary strengths of both models and provides insights to enhance cryptocurrency price prediction in dynamic markets such as Indonesia.
5:10 PM – 5:50 PM Keynote
The Promise and Peril of Generative AI
Dr. Mark Maybury, Vice President, Commercialization, Technology & Strategic Innovation for Lockheed Martin
Dr. Mark Maybury is the vice president, Commercialization, Engineering & Technology for Lockheed Martin, responsible for leading efforts to commercialize dual-use products and services across the corporation. Dr. Maybury’s prior roles include first chief technology officer for Stanley Black & Decker, Chief Scientist of the U.S. Air Force, Chief Technology and Chief Security Officer at MITRE and Director of the National Cybersecurity FFRDC. Read more >>
5:50 PM – 6:00 PM Awards and Closing
7:30 AM – 6:00 PM Exhibitors – All Day
Created and maintained by Ballos Associates