NCTA 2025 Abstracts

Full Papers

Paper Nr:	21
Title:	Determining Optimal Pixel Resolution for Object Detection in Satellite Imagery: A Class-Specific Approach
Authors:	Daniel C. Fox, John Prominski, Amit Virchandbhai Prajapati, Kendall Haddigan, Gabriel Barbosa, Shubham Dashrath Wagh, Adam Nolan, Daniel Zwillinger, Chun-Kit Ngan, Fatemeh Emdad and Elke Rundensteiner
Abstract:	This paper presents a comprehensive exploration of determining optimal pixel resolutions for object detection in satellite imagery through a class-specific approach. Object detection in satellite imagery, critical for applications such as urban planning, environmental monitoring, and military surveillance, poses unique challenges due to high image resolutions, small object sizes, and computational demands. We propose a reusable pipeline de-signed to automate the discovery of the "knee point" on resolution-performance curves, achieving a balance between detection accuracy and computational efficiency. The pipeline integrates modules for data preprocessing, model fine-tuning, performance evaluation, and automated report generation. Utilizing the xView1 dataset and the YOLOv8 object detection model, we systematically analyze resolution images across 48 moveable object classes. Our findings show that lower-resolution images can yield competitive performance, significantly reducing resource demands, especially among object classes that perform well at high resolutions. This bridges existing research gaps while emphasizing modularity, efficiency, and usability. On average, across all object classes considered in this pipeline run the designated knee point presented a 57% reduction in pixel data with an 80% retention of the highest detection performance achieved at any resolution. Further, if we narrow our scope down to the top 15 performing classes, we find that the designated knee point presents a 71% reduction in pixel data, enabling a ground coverage area 3.4 times larger than what is achievable at the highest resolution, while retaining 88% of the detection performance and maintaining the same image dimensions and hardware capabilities.

Paper Nr:	61
Title:	Innovative Techniques for Efficient Hyperdimensional Computing on Hardware: Enhance Accuracy and On-Fly Hypervector Generation
Authors:	Saeid Jamili, Sabereh Taghdisi Rastkar, Marco Angioli and Mauro Olivieri
Abstract:	Hyperdimensional Computing (HDC) encodes data and learning operations into high-dimensional vectors (hypervectors), enabling robust and rapidly adaptable machine learning on resource-limited platforms such as Field-Programmable Gate Arrays (FPGAs) and edge devices. Despite this potential, conventional HDC systems often demand extensive memory resources to store base, level, and class hypervectors, which can limit scalability and performance in hardware implementations for Artificial Intelligence (AI) and Internet of Things (IoT) applications. This paper addresses these issues through two main innovations. First, it introduces a combinational-logic approach that generates hypervectors on the fly, eliminating the need for large lookup tables and thereby substantially reducing memory overhead. Second, it presents an orthogonal hypervector generation scheme based on sequences such as Hadamard, Walsh, and Gold, ensuring highly uncorrelated representations that enhance classification accuracy (particularly in single-shot learning) while remaining effective over multiple training epochs. Experimental evaluations on standard benchmarks, including ISOLET and UCI-HAR, show notable gains in classification performance as well as significant reductions in memory consumption and lookup table usage. These results highlight the viability of integrating logic-based hypervector synthesis with orthogonal vector design to create an efficient, power-conscious, and high-throughput HDC framework suitable for real-time edge AI and IoT scenarios. By uniting these techniques, the proposed approach advances the practical deployment of hyperdimensional computing in embedded and resource-constrained environments.

Paper Nr:	68
Title:	A Structured Survey of Anomaly Types and Classification-Based Detection Models in IoT
Authors:	Atefeh Gilvari, Ziad Kobti, Narayan Kar, Nasrin Tavakoli and Rajeev Verma
Abstract:	In dynamic Internet of Things (IoT) environments, traditional anomaly detection surveys often treat all anomalies as a unified concept, overlooking the distinct characteristics posed by specific anomaly types. This paper presents a structured survey and comparative analysis of anomaly detection models, organized by type of anomalies such as drift, novelty, bias, noise, constant-value, and stuck-at-zero anomalies. Each anomaly type is formally defined along with its theoretical foundation, followed by a systematic review and analysis of how model effectiveness varies across these types to identify techniques best suited for each. Our findings emphasize the need for interpretable, adaptive, and type-aware anomaly detection systems and outline open challenges in unified benchmarking, cross-type detectors, and ontology development for anomaly classification. A novel contribution of this work is a broader mapping framework that illustrates how various models differ in their ability to detect specific anomaly types across different IoT domains, offering insight into the generality and specialization of current detection approaches.

Paper Nr:	98
Title:	MLP Model for Prediction of Pellet Combustion: How to Deal with Small Datasets
Authors:	Philippe Thomas, Eliott Gauthey Franet, Yinling Liu, Hind El Haouzi, Jérémy Hugues Dits Ciles and Yann Rogaume
Abstract:	The increase in the use of wood in general, and pellets in particular, for individual heating requires attention to be paid to optimizing the combustion of these pellets, particularly in terms of pollutant emissions such as carbon monoxide (CO). This quest for optimization is complicated by the intrinsic variability of wood and, therefore, pellets. In this context, this paper proposes building a neural network model to predict the CO content in smoke based on pellet characteristics and stoves characteristics and settings. However, experiments are costly and time-consuming, which limits the size of the available dataset. In this context, we propose a methodology aimed at training a multilayer perceptron using a small dataset while reducing the risk of overfitting. This methodology is based on finding the minimal network structure and using a robust learning algorithm. The results show that the robust algorithm effectively limits the risk of overfitting and that the final model retains its generalization capabilities despite the small size of the dataset.

Paper Nr:	145
Title:	Towards Robust Urban Parking Violation Prediction Using Graph Kolmogorov–Arnold Networks and Liquid Neural Networks
Authors:	Mohammad Reza Mohebbi, Javad Mohebbi Najm Abad, Elahe Kafash and Mario Döller
Abstract:	Illegal parking in urban environments disrupts traffic flow, causes greenhouse gas emissions, and poses a threat to pedestrians and cyclists. Traditional Intelligent Transportation Systems (ITS) are based on high-cost surveillance and video analysis that typically does not take into account the dynamics and complexity of the urban environment. To address these limitations, this study overcomes this gap by proposing an intelligent parking violation prediction framework using a hybrid Spatio-temporal Graph Neural Network (STGNN) approach, which combines Graph Kolmogorov-Arnold Networks (GKAN) and Liquid Neural Networks (LNN). The GKAN model excels at uncovering intricate spatial patterns in the urban dataset, while the LNN model has intrinsic dynamic temporal variations in real time due to its adaptive learning capability. This integration of spatio-temporal relationships of metropolitan datasets is effective modeling and thus can be robust across diverse urban scenarios. The proposed approach is well-suited for practical usage in real-world applications and achieves a high prediction accuracy with a high R² score of 0.95, and shows significant improvement in other metrics such as MAE and MSE. These results underscore the performance of the proposed GKAN-LNN framework in addressing the challenges presented by the parking violation prediction task to develop safe, sustainable, and well-governed urban settings.

Short Papers

Paper Nr:	38
Title:	Re-Ranked Transformer: New Strategy Based on Misspellings and Typos Pattern Analysis for Keystroke Biometrics Improvement
Authors:	Nabila Mansouri, Salwa Sahnoun, Hedi Fekhi, Ahmed Ben Ali and Ahmed Fakhfakh
Abstract:	Person identification using keystroke biometrics offers a scalable solution for identity verification in behavioral biometrics. This study introduces a framework where a transformer model serves as the baseline for capturing complex spatio/temporal patterns in keystroke features. To enhance accuracy, the model’s output is re-ranked using k-reciprocal nearest neighbors (k-RNN), which encodes neighborhood relationships into a feature vector for re-ranking under the Jaccard distance. The proposed method integrates misspellings and typos patterns, particularly using backspace key for correction, as a weighting factor to refine the final identification distance. This piepline, combining the transformer baseline with k-RNN re-ranking and typing error adjustments, demonstrates significant improvements in person identification. Experimental results on Aalto keystroke database achive and EER about of 1.60. These findings validate the effectiveness of our proposed method and highlight its potential for secure and non-invasive applications.

Paper Nr:	59
Title:	Degradation-Aware Energy Management in Residential Microgrids: A Reinforcement Learning Framework
Authors:	Danial Zendehdel, Gianluca Ferro, Enrico De Santis and Antonello Rizzi
Abstract:	This paper presents a degradation-aware reinforcement learning (RL) framework for real-time energy management in residential microgrids, focusing on optimizing lithium-ion battery usage while balancing economic benefits and battery longevity. We employ the Soft Actor-Critic (SAC) algorithm, implemented via Stable Baselines3, to learn non-linear dispatch policies for a 5.2 kWh LiCoO$_2$ battery pack, with degradation modeled using a simplified energy-throughput approach calibrated with NASA dataset measurements. The framework is tested across diverse household profiles over 1-year and 10-year simulations. Results show that RL-SAC outperforms a Model Predictive Control (MPC) baseline, extending battery life and reducing energy purchases in both simulations. These findings highlight RL-SAC’s potential for practical deployment in microgrids, offering a scalable solution for sustainable energy management.

Paper Nr:	63
Title:	A Universal Urban Electricity-Demand Simulator for Developing and Evaluating Load-Scheduling and Forecasting Systems
Authors:	Sabereh Taghdisi Rastkar, Saeid Jamili, Enrico De Santis and Antonello Rizzi
Abstract:	Accurately modeling city-scale electricity demand is fundamental to the design of data-driven forecasting, demand-response (DR), and grid-control solutions. This paper introduces an open-source simulator that recreates city-wide electricity demand by linking four key models in one sub-hourly loop: (1) a weather generator that produces temperature, solar, wind, humidity, and price signals; (2) a building-load model that converts these time series into residential, commercial, and industrial demand; (3) a demand-response (DR) aggregator that trims or shifts loads when prices are high or feeder limits are reached; and (4) a distribution-network power-flow solver that checks voltages and line ratings on a feeder. To this aim we adopted the pandapower tool. All outputs like weather, sector loads, curtailed demand, and grid states are streamed to disk in compressed chunks, so multi-year studies run without exhausting memory. A built-in test suite strengthens scientific rigour: it (i) verifies bit-for-bit reproducibility under fixed random seeds, (ii) confirms that temperature and wind residuals follow the intended Normal and Weibull laws, (iii) uses Ljung–Box tests to show the expected time-correlation, (iv) computes steady-state confidence intervals with the batch-means method, and (v) checks model sensitivity by sweeping key parameters. Together, these features give planners, utilities, and researchers a reliable tool for exploring how weather, occupant behaviour, DR rules, and network limits interact.

Paper Nr:	66
Title:	Drowsiness Detection with Time-Series Classification Using HRV Features
Authors:	Duarte Valente, Artur Ferreira and André Lourenço
Abstract:	Drowsy driving significantly increases the risk of road accidents and crashes. However, the drowsy state remains difficult to detect in real-time. This study presents a supervised learning approach to driver drowsiness detection using heart rate variability (HRV) features derived from electrocardiogram (ECG) signals. Data was collected in a driving simulator from participants under different levels of sleep deprivation, with subjective sleepiness levels recorded using the Karolinska Sleepiness Scale (KSS). HRV features were extracted in both time and frequency domains and used to train classification models, including neural networks and long short-term memory (LSTM) architectures. The experimental results show that HRV-based features can effectively model drowsiness levels, with LSTM models outperforming common baseline methods. Our approach copes with the progressive nature of the drowsiness state and contributes to the development of intelligent in-vehicle systems capable of non-invasively monitoring driver alertness and issuing timely alerts to prevent fatigue-related accidents.

Paper Nr:	74
Title:	Fine-Tuning Prototypes for Cross-Domain Few-Shot Image Classification Using Contrastive Objective
Authors:	Abhishek Mahajan, Ziad Kobti and Bishwadeep Sikder
Abstract:	Cross-domain few-shot learning (CFC) seeks to enable accurate classification in novel visual domains using only a few labeled samples per class. However, it remains a challenging task due to prototype misalignment and domain-induced embedding distortions, in addition to lack of purified data. While the state-of-the-art transformation network effectively realigns prototypes in one-shot settings, it is not designed for broader domain generalization or transformer-based architectures in cross-domain settings. We introduce CosRestViT, a prototype-aware few-shot learning framework designed to tackle the problems of scarce data with a lot of outliers. Our method integrates a custom transformation network with a Vision Transformer (ViT) backbone to recalibrate noisy embeddings and align them with class-level prototypes. To enhance semantic consistency in the embedding space, we fine-tune the model using a hybrid contrastive objective that combines cross-entropy loss with CoSENT’s ranking loss. We pretrain CosRestViT on the large-scale base dataset using a prototype-aware contrastive loss to enhance the representational power of ViT and ensure effective clustering of semantically similar samples. The model is then evaluated on the Meta-Dataset benchmark under two training regimes: (i) training on all seen datasets, and (ii) training exclusively on ImageNet for transfer learning. Experimental results demonstrate that CosRestViT outperforms existing baselines across both seen and unseen domains, achieving superior generalization in low-data scenarios.

Paper Nr:	82
Title:	Dataset-Independent Approach for Generating Synthetic Data in Optical Defect Detection
Authors:	Christian Linder, Steffen Geinitz and Sebastian Maier
Abstract:	Deep learning techniques have become increasingly important in the field of optical defect detection in recent years, often outperforming traditional image processing methods. However, the performance of deep learning methods is highly dependent on the amount of training data, which poses a challenge in industrial settings where labelled defect images are often scarce. This paper presents a novel dataset-independent approach for generating synthetic defect images that improves the performance of deep learning models while requiring minimal real-world data. Our methodology uses a pix2pix model to generate realistic defect images. The pix2pix model is trained on defect images and images where the defects are segmented out. After training, the model is applied to defect-free images to transform them into defect images using random segmentation masks. This procedure ensures that the defect location is preserved and that the synthetic data can also be used for segmentation applications. The approach is validated on an open source dataset. The results show that the proposed synthetic data generation approach reduces class imbalance and leads to improvements in model accuracy and recall.

Paper Nr:	91
Title:	Combining Large-Scale and Domain-Specific Datasets for Hate Speech Severity Modeling: A Regression-Based Approach
Authors:	Andrew Asante and Petr Hajek
Abstract:	This study proposes a regression-based framework for modeling the severity of online hate speech, addressing the limitations of traditional classification approaches that overlook the nuanced and continuous nature of harmful language. Leveraging a compiled dataset of over 3.1 million English-language social media posts from Reddit, Twitter, and Wikipedia, we fine-tune both general-purpose and hate speech-specific transformer models to predict real-valued severity scores. Our findings show that domain-specific models, particularly HateBERT, outperform general-purpose alternatives in capturing subtle gradients of hatefulness. The proposed approach enables more context-aware and proportionate content moderation, while also highlighting challenges related to annotation subjectivity, lack of conversational context, and cross-platform generalization. This work advances the field by demonstrating the feasibility and utility of severity estimation as a scalable alternative to categorical hate speech detection.

Paper Nr:	126
Title:	Multi-Subspace SVD Generators for Continual Learning
Authors:	Christiaan Lamers, Ahmed Nabil Belbachir, Thomas Bäck and Niki van Stein
Abstract:	Catastrophic forgetting is a persistent obstacle for continual learning on memory-constrained edge devices. We introduce Multi-Subspace SVD Generators (MSSG), an extension of Lightweight SVD Generators that replaces each single global SVD with a small ensemble of low-rank SVDs per task and class. Each subspace captures a local linear patch; together they approximate the non-linear data manifold while keeping storage cost proportional to the rank, not the number of replay samples. A closed-form memory model allows MSSG to trade subspace count and rank against an equivalent raw-sample buffer, enabling fair comparison to experience-replay baselines. Across five image benchmarks; MNIST, Fashion MNIST, NOT MNIST, CIFAR10 and Tiny ImageNet, MSSG (i) outperforms its single-generator predecessor, (ii) matches or exceeds the accuracy of Experience Replay with a large buffer, and (iii) does so with less than one-tenth of their memory footprint. Because MSSG stores only compact factorised statistics, both rehearsal and generator updates run in milliseconds on resource-limited hardware, making it a practical drop-in replacement for replay buffers in on-device lifelong learning applications.

Paper Nr:	143
Title:	From High-Frequency Sensors to Noon Reports: Using Transfer Learning for Shaft Power Prediction in Maritime
Authors:	Akriti Sharma, Dogan Altan, Dusica Marijan and Arnbjørn Maressa
Abstract:	With the growth of global maritime transportation, energy optimization has become crucial for reducing costs and ensuring operational efficiency. Shaft power is the mechanical power transmitted from the engine to the shaft and directly impacts fuel consumption, making its accurate prediction a paramount step in optimizing vessel performance. Power consumption is highly correlated with ship parameters such as speed and shaft rotation per minute, as well as weather and sea conditions. Frequent access to this operational data can improve prediction accuracy. However, obtaining high-quality sensor data is often infeasible and costly, making alternative sources such as noon reports a viable option. In this paper, we propose a transfer learning-based approach for predicting vessels' shaft power, where a model is initially trained on high-frequency data from a vessel and then fine-tuned with low-frequency daily noon reports from other vessels. We tested our approach on sister vessels (identical dimensions and configurations), a similar vessel (slightly larger with a different engine), and a different vessel (distinct dimensions and configurations). The experiments showed that the mean absolute percentage error decreased by 10.6% for sister vessels, 3.6% for a similar vessel, and 5.3% for a different vessel, compared to the model trained solely on noon report data.

Paper Nr:	40
Title:	Towards Generalizing Deep Reinforcement Learning Algorithms for Real World Applications
Authors:	Daniel Ruiru, Nicolas Jouandeau and Dickson Owuor
Abstract:	Generalization is a major problem in reinforcement learning (RL), as agents struggle in environments outside their training sets. Often caused by overfitting during the training phase, this issue limits the application of RL in the real world. This paper tries to solve the generalization problem by using a domain randomization technique during the training period. Using two real-world problems; the financial market and agriculture (crop production), this paper trains classical deep reinforcement learning algorithms (DQN and PPO) in different and randomized environments (MDPs). Agents trained in randomized environments generalize better than those trained in single environments (baseline agents). This conclusion is based on the results, in which the agents trained in the randomized environments achieve higher cumulative rewards.

Paper Nr:	70
Title:	Assessing Driving Style from Telematics Data with a Two-Stage Clustering Approach
Authors:	Duarte Valente, Luís Loureiro, Artur Ferreira and André Lourenço
Abstract:	The way a person drives is a relevant source of information to make decisions about that person, in some contexts. For instance, insurance companies set vehicle insurance fees as functions of static variables, such as the age of the driver, the number of years one holds a driving license, and the driving history. These variables may not reflect the everyday behavior of the driver on the road, thus ending up by penalizing good drivers that are young. Another example if the fleet management task, in which it is relevant to know who the best drivers are, to make the best trip planning decisions. In this paper, we follow a pay-as-you-drive approach, to devise a driver style identification approach, based on driver behavior data. Using anonymous data records with the trips from different drivers, we build a dataset and we apply unsupervised dimensionality reduction and clustering techniques. The experimental results show clusters with distinct trip styles. Many drivers show a non-aggressive driving style, some have an aggressive style and a few have a risky style.

Paper Nr:	77
Title:	OS-QLR: One-Shot Quantized Latent Refinement for Fast and Efficient Image Generation
Authors:	Peng Li, Roman Senkerik and Adam Viktorin
Abstract:	This article introduces One-Shot Quantized Latent Refinement (OS-QLR), an original two-stage generative framework designed for both high-quality image generation and enhanced computational efficiency. OS-QLR first learns a compact, discrete latent representation via a Vector Quantized Variational Autoencoder (VQ-VAE), then employs a single-step refinement network in this latent space to generate clean, plausible samples from noisy or random inputs. Experimental results on FashionMNIST and CIFAR-10 datasets demonstrate that OS-QLR consistently yields superior image quality, characterized by sharper details, fewer artifacts, and significantly lower Fréchet Inception Distance scores compared to unrefined VQ-VAE. Furthermore, OS-QLR exhibits robust performance against various levels of latent space corruption. Critically, its training process is reduced from days or even weeks to hours compared with the Diffusion Models, Generative Adversarial Networks (GANs) and Autoregressive image generation models, its non-iterative sampling enables rapid image generation, positioning OS-QLR as a compelling and efficient alternative to current computationally intensive generative models.

Paper Nr:	152
Title:	Data Augmentation for Neuroaesthetics Analysis
Authors:	Maurizio Palmieri, Marco Avvenuti, Francesco Marcelloni and Alessio Vecchio
Abstract:	Neuroaesthetics investigates the neural activities during aesthetic experiences, using EEG recordings or fMRI images to decode the perception of visual art. However, studies in this domain are hindered by the limited and imbalanced nature of datasets, which is due to the subjective and resource-intensive nature of data collection. This study examines the effectiveness of various data augmentation strategies in enhancing EEG classification performance for neuroaesthetic analysis. We experiment three different EEG augmentation techniques, namely Signal Segmentation and Recombination, Temporal and Spatial Reconstruction Data Augmentation, and Gaussian Noise Addition. Furthermore, once extracted the features from the signals, we applied an instance-level data augmentation algorithm, namely SMOTE. We tested the four augmentation techniques individually, as well as SMOTE applied in cascade with the three EEG-specific augmentation methods, using stratified ten-fold cross-validation and leave-one-subject-out validation strategies. Results show that Gaussian Noise Addition, particularly when combined with SMOTE for generalization, yields consistent performance improvements in both accuracy and F-score. Conversely, Temporal and Spatial Reconstruction Data Augmentation often degrades classification performance (up to -9.95% of accuracy). Finally, Signal Segmentation and Recombination achieved the best improvement in the leave-one-subject-out analysis (+2.2% accuracy). Our findings present that appropriate data augmentation can enhance model generalization for aesthetic experience classification.