NCTA 2026 Abstracts

Full Papers

Paper Nr:	32
Title:	Internal Information Control for Interpreting Neural Networks
Authors:	Ryotaro Kamimura
Abstract:	The present paper aims to show how to control mutual information internally without influences from outside, although this setting is purely idealized. This internal approach is proposed to overcome the limitations of the conventional external approach, which focuses mainly on information originating from outside the network. In the internal approach, it is assumed that many different levels exist for obtaining information. In this paper, the forward and backward levels are assumed to compete with each other during learning. By controlling this competition, we attempt to interpret how network configurations are internally produced and what they try to achieve through mutual information augmentation. The method was applied to the analysis of French verse and prose because we assume that the complicated devices used in verse aim to increase information against the natural tendency of information decay. The findings can be summarized in three points. First, in the internal approach, the forward and backward mutual information measures were effective in increasing the actual forward and backward information, respectively. In particular, the backward method could produce hidden layers with greater information content even when they were far from the base layers. Second, internal mutual information maximization was not necessarily related to improved generalization. For improving generalization, the ratio of backward to forward information should be as close to equal as possible, meaning that generalization can be improved only in this harmonic state, where the mutual information of the two levels is balanced. Finally, regarding the meaning of the constraints imposed in verse, they attempt to increase information even between characters that are far from the base character. Namely, they try to cope with the problem of natural information decay by introducing specific devices into natural language.

Paper Nr:	42
Title:	Stimulus-Aware Continuous Valence Prediction from EEG and Video via Mixture-of-Experts
Authors:	Ali Amini, Mohammad Alijanpour, Hamed Mirzabagherian, Ameer Hamza, Rytis Maskeliūnas and Robertas Damaševičius
Abstract:	Valence during media consumption is shaped by both the external stimulus and the viewer’s internal neural response. We model this setting after MAHNOB-HCI by coupling an EEG encoder with a stimulus-side video representation to enable continuous valence regression. The EEG branch employs a hierarchical spatiotemporal design that summarizes temporally salient activity for each electrode and then captures cross-electrode interactions via attention, yielding a compact embedding. In parallel, we extract clip-level visual descriptors from a frozen VideoMAE v2 backbone and map them into a shared latent space. A lightweight mixture-of-experts regressor combines modality-specific predictions through an input-conditioned gate, enabling adaptive reliance on EEG or video across windows. Under trial-grouped 5-fold cross-validation, the proposed EEG+Video model achieves PCC=0.795, CCC=0.766, and RMSE=0.026, outperforming representative prior results under comparable protocols. Beyond accuracy, we analyze the EEG embedding geometry and observe an approximately monotonic increase in valence-label differences with latent distances, suggesting that the learned representation is structured consistently with the continuous valence continuum.

Paper Nr:	57
Title:	Localising Drone RF Emissions in Spectrograms: Automatic Annotation and Noise-Robust Object Detection
Authors:	Stefan Glüge, Julian Kraft, Christof Schüpbach and Matthias Nyfeler
Abstract:	Reliably detecting unmanned aerial vehicles (UAVs) has become a priority for safety systems around critical infrastructure. The passive operation and long detection range of radio frequency (RF)-based detection make it attractive. We applied a simple object detection model to real RF drone recordings, treating each signal burst as an object. To this end, we created an automatic annotation procedure that generated bounding boxes for drone signal bursts using energy-based segmentation. We conducted a noise-robustness analysis on the resulting dataset, showing that progressive noise augmentation with Gaussian noise during training yields strong detection performance, with a mean average precision (mAP)50 ≈ 0.95 down to a signal-to-noise ratio (SNR) of −20 dB, without degrading clean-data performance. The detector recovered the location of each burst in time and frequency, revealing parameters such as center frequency, bandwidth, and dwell time. Furthermore, the inference time for a single spectrogram was ≈3 ms, which is fast enough for real-time operation. Although robustness toward Gaussian noise is a necessary but not sufficient condition for real-world applicability, we believe our findings can be used to develop better models for robust UAV detection systems.

Paper Nr:	60
Title:	A Reinforcement Learning Model for Microgrid Energy Management under Uncertainty and Battery Degradation
Authors:	Gianluca Ferro, Alessio Orlandi, Mohammad Ali Sherafat, Enrico De Santis and Antonello Rizzi
Abstract:	Residential microgrids with photovoltaic generation and Battery Energy Storage Systems require control strategies capable of reducing energy costs while limiting battery degradation under uncertain operating conditions. This work presents a reinforcement learning Energy Management System based on Proximal Policy Optimization, where battery wear is incorporated into the reward through a degradation cost related to depth of discharge. The controller learns charge and discharge policies from interaction with the microgrid environment while using forecast information on future load and photovoltaic generation. The evaluation involves a residential case study and compares the proposed controller with Rule-Based Control and Model Predictive Control. The results show that the learned controller offers a more favorable compromise between grid cost and battery degradation than the greedy benchmark and remains competitive with the predictive baselines. In the Pareto comparison, longer forecast horizons improve the ability of the policy to reduce grid operating cost while limiting battery wear, and several configurations attain lower total cost than the corresponding oracle Model Predictive Control cases.

Paper Nr:	61
Title:	Artemix-ViT: Design and Implementation of a Compact Vision Transformer Accelerator on FPGA
Authors:	Saeid Jamili, Sabereh Taghdisirastkar, Stefano Perticaroli, Abdallah Cheikh, Mhd Rashed Al Koutayni and Mohamed Selim
Abstract:	This paper presents a compact FPGA accelerator for edge image classification based on a hardware-aware Vision Transformer (ViT). The deployed network maps 32 × 32 RGB inputs to 65 tokens of width 80 and uses four encoder blocks, two attention heads, RMSNorm, and a patch-local depthwise branch. Focused design-space exploration and teacher-student distillation produce the student model; export then converts the trained weights, gains, shifts, filters, class-token data, and positional constants into a fixed-point parameter package. The hardware flow has two implementations. V1 is a modular HLS prototype for checking tensor ordering, package decoding, and arithmetic. V2 is the final RTL design: a CSR shell starts a descriptor sequencer, a parameter service, banked activation storage, and reusable dense, online-attention, RMSNorm, vector/layout, and depthwise engines. One host launch runs the graph through a 77- step schedule while intermediate int16 tensors remain on chip. V2 closes timing with positive post-implementation slack, runs at a measured board clock of 92.3 MHz, and completes inference in about 4.51M cycles/image, corresponding to 48.9 ms/image and 20.5 images/s with parameter service included. Under the release 10,000-image CIFAR-10 board-evaluation protocol, the final board package reports 85% accuracy.

Paper Nr:	72
Title:	RelFG-Net: Dynamic Prototype-Guided Graph Refinement for Accurate Fine-Grained Visual Classification
Authors:	Akkhar Ulok, Atefeh Gilvari, Ziad Kobti and Sudhir Paul
Abstract:	Fine-grained visual classification is challenging due to subtle interclass differences and strong visual similarity among categories. We propose RelFG-Net, a relational fine-grained framework that combines graph reasoning with convolutional feature learning to model inter-class relationships. During training, dynamic class prototypes are constructed and refined through a sparsified class-relation graph using graph convolution layers, enabling structured message passing among similar categories. Experiments on CUB-200-2011, FGVCAircraft, and Stanford Cars demonstrate consistent improvements over DTRG. RelFG-Net achieves 2.35% Top-1 improvement on Aircraft, 1.34% on CUB, and 2.29% on Cars, while also improving Top-5 accuracy and F1-score across datasets on the same backbone. These results show that learnable relational reasoning over evolving class representations effectively improves fine-grained recognition. The full implementation is available on GitHub (github.com/Akkhar1/RELFG-Net).

Paper Nr:	95
Title:	RAG4Jail: Boostrapped Retrieval-Augmented Jailbreak Generation for Large Language Models
Authors:	Miro Confalone, Mattia Cervelli, Emilio Coppa, Irene Finocchi and Alessio Martino
Abstract:	Jailbreaking, that is, crafting prompts that bypass the safety constraints of Large Language Models (LLMs), remains a critical threat to deployed AI systems. We present RAG4JAIL, a bootstrapped retrieval-augmented pipeline for automated jailbreak generation. Given a seed query, RAG4JAIL retrieves semantically related examples from a vector store of community-reported attacks and feeds them as adversarial context to a generator LLM, which synthesizes a novel candidate under one of ten established jailbreaking techniques. A key design element is bootstrapped retrieval: after each round, successful attacks are reinjected into the corpus, progressively refining the pool available to subsequent rounds. We instantiate RAG4JAIL in a goal-driven variant, supporting human-in-the-loop red-teaming, and an attack-driven variant, for systematic strengthening of existing benchmark suites. Experiments across four safety-aligned target models and six public datasets show that even a single bootstrapping round more than doubles attack success rates on the most robust targets, with further rounds yielding consistent but diminishing gains. The improvement is technique-agnostic: all ten jailbreaking techniques benefit from retrieval and bootstrapping, implying that defenses targeting any specific attack pattern offer only partial protection. RAG4JAIL matches or outperforms the best comparable black-box baselines without requiring gradient access, model internals, or target-specific tuning.

Paper Nr:	102
Title:	MSHGNN: Metapath Selection for Heterogeneous Graph Neural Network
Authors:	Johan Le Boursicaud, David Gross-Amblard and Zoltan Miklos
Abstract:	Heterogeneous Graph Neural Networks (HGNNs) have become a widely adopted framework for modeling relational data with multiple node and edge types. Among existing approaches, metapathbased methods are particularly effective, capturing high-order semantic relations while remaining interpretable. Yet selecting an appropriate metapath set is a persistent challenge: overly large sets inflate computational cost and introduce semantic noise, while overly restrictive choices risk excluding informative long-range dependencies. We propose MSHGNN (Metapath Selection for Heterogeneous Graph Neural Network), a two-stage framework for node classification that addresses this trade-off directly. In the first stage, a lightweight training phase estimates the relative importance of candidate metapaths via semantic attention. In the second stage, the model is retrained using only the top-K most informative metapaths, with an optional semantic diversity constraint to reduce redundancy among the selected paths. Extensive experiments on standard heterogeneous graph benchmarks demonstrate that MSHGNN achieves competitive or superior accuracy compared to stateof-the-art HGNNs, while substantially reducing computational complexity and training time.

Paper Nr:	176
Title:	Distilling Full-Reference Image Quality Metrics into Neural No-Reference Surrogates
Authors:	Leonardo Sgroi, Mattia Gatti, Ignazio Gallo and Silvia Corchs
Abstract:	Full-reference (FR) image quality assessment metrics such as SSIM, FSIM, and IW-SSIM provide reliable and interpretable measures of perceptual similarity. Their applicability is nevertheless limited by the requirement for a pristine reference image and by their poor suitability for large-scale, real-time, and differentiable optimization pipelines. This work introduces a deep learning framework that learns neural surrogates of FR metrics, reproducing their values from a single degraded image. Training relies on synthetically distorted images, for which groundtruth FR values are computed against pristine references, so that SSIM, FSIM, IW-SSIM, and SR-SIM can be regressed in a reference-free setting. A controlled comparison of convolutional and transformer-based backbones shows that compact transformer-based backbones attain the highest surrogate fidelity over the full score range, where MobileViT-v2 reaches R 2 = 0.8871 for SSIM and R 2 = 0.8536 for FSIM on a large-scale ImageNet-based test set. Restricting the evaluation to high-quality intervals exposes the limits imposed by the compressed dynamic range of the targets, and does so very unevenly across metrics: accuracy collapses for SSIM, degrades moderately for IW-SSIM and FSIM, and is unaffected for SR-SIM, while the ranking of the backbones itself changes. A further evaluation on TID2013 shows that the learned surrogates retain a measurable correlation with mean opinion scores, despite the absence of subjective annotations during training.

Short Papers

Paper Nr:	13
Title:	Transformer Infused Deep Reinforcement Learning for Resource Allocation in Crop Production
Authors:	Daniel Kiguru Ruiru, Nicolas Nicolas and Dickson Owuor
Abstract:	The agricultural sector faces serious challenges in meeting the needs of food security. Crop management is one of these challenges, where a balance between productivity and environment impact is needed. Such a complex issue is seen in fertilizer application as crop management practices try to increase crop yields while reducing environmental impacts. This issue represent a classical case of resource allocation problem where machine learning can be applied. Predictive methods like Deep Reinforcement Learning (DRL) are gaining popularity in solving such problems, which this paper uses to sustainably apply fertilizers to crops. This paper presents a fairly new solution to this problem by combing transformers with DRL agents. These transformer-based DRL agents are trained in the same environment as baseline DRL agents to compare their performance. Transformerbased agents show considerable superiority in balancing the needs of crop production and environmental sustainability. Ultimately, transformer-based agents get better yields and cumulative reward without applying excessive amounts of nitrogen fertilizer.

Paper Nr:	18
Title:	Comparison of Recognition Capabilities in Neural Networks and in Birds on Different Types of Landscapes
Authors:	Margarita Zaleshina and Alexander Zaleshin
Abstract:	The need to reduce recognition errors while accounting for changing environments through learning exists in both neural networks and in nature. In our study, we examined general patterns in recognizing spatial data in natural and urban landscapes for neural networks and flying pigeons. We tested the number of errors made by both neural networks and birds during recognition in the same 180-square-kilometer area. For neural networks, errors were defined as incorrect identification of an object type, while for pigeons, errors were defined as significant deviations from the initial route. Using heatmaps of the distribution of recognition errors across the study area, we identified the locations and types of areas that posed recognition difficulties. This resulted in the formation of localized clusters consisting of spatial elements in which recognition accuracy varied significantly. It was found that these clusters belong to different types of terrain (rural, urban, and mixed). The results of our study can be used to select the settings for neural networks used in the analysis of remote sensing datasets.

Paper Nr:	27
Title:	Cost-Aware Embedding Dimension Selection via Accuracy and Training-Cost Surrogate Models
Authors:	Takahisa Imagawa, Aoto Uchida and Shuichi Enokida
Abstract:	Transformer-based models are widely used in practical deep learning applications, but their deployment requires balancing accuracy, model-selection cost, and inference cost. The embedding dimension is a key structural hyperparameter in Transformer-based models because it affects representational capacity and training time and serves as a proxy for inference FLOPs. However, evaluating many candidate embedding dimensions by training each of them for a sufficient number of epochs is computationally expensive. In this study, we propose a novel method for selecting the embedding dimension while taking training cost into account. The proposed method combines a validation-accuracy prediction model and a training-cost model, both of which take the embedding dimension and the number of training epochs as inputs. Our validation-accuracy prediction model uses a monotone, saturating form with only a few parameters as a parsimonious inductive bias, keeping the regression problem low-dimensional and enabling repeated fitting from a small number of training observations. Experimental results using Action Transformer and the Driver Monitoring Dataset show that the proposed method selects smaller embedding dimensions on average while achieving higher validation accuracy than comparison methods under comparable total training-cost budgets.

Paper Nr:	31
Title:	AD Early Screening Method Based on Dynamic Graph Attention Network and Multi-Scale Entropy
Authors:	Troy William Deng and Xuanming Shi
Abstract:	Early detection of Alzheimer’s Disease (AD) is critical for effective intervention. Electroencephalography (EEG) provides a non-invasive and costeffective approach for monitoring brain activity; however, existing methods often overlook temporal dynamics and rely on simplified handcrafted features, limiting their ability to characterize nonlinear neural activity. To address these limitations, we propose a Dynamic Graph Attention Network with Multi-scale Entropy (DGAT-MSE) for EEG-based early AD screening. Specifically, multi-scale entropy (MSE) is employed to quantify nonlinear dynamics of EEG channels within sliding temporal windows, while phase-locking value (PLV) is used to construct dynamic functional connectivity graphs. The proposed D-GAT framework jointly captures spatial brain network topology and temporal variations in EEG activity. Furthermore, GNNExplainer is introduced to enhance model interpretability by identifying critical brain regions and informative temporal windows. Experiments conducted on the OpenNeuro dataset demonstrate that the proposed method achieves a classification accuracy of 85.71%, outperforming conventional approaches based on static connectivity or power-spectrum features. Ablation studies further verify the effectiveness of both MSE features and dynamic graph modeling. Interpretability analysis reveals that frontopolar (Fp2, Fp1) and frontal midline (Fz) regions contribute most significantly to AD classification during resting-state EEG, followed by posterior temporal (T6) and occipital (O1) regions. This frontal-lobe dominance aligns with recent evidence that prefrontal EEG activity serves as a cognitive reserve marker in AD and that frontal functional connectivity undergoes compensatory enhancement in early-stage neurodegeneration. These findings offer a novel interpretable framework for early AD screening and demonstrate the potential of extending EEG-based graph learning to other neurological disorders.

Paper Nr:	52
Title:	Enhancing Prediction Performance in Time Series Forecasting with Anomaly Signals Detected by Hyperdimensional Computing
Authors:	Sabereh Taghdisi Rastkar, Saeid Jamili, Enrico De Santis and Antonello Rizzi
Abstract:	Accurate electricity-demand forecasting is critical for reliable and sustainable smart-grid operation, yet irregular events and data anomalies can substantially degrade predictive performance, particularly during high-demand periods. This paper proposes an anomaly-aware forecasting framework that augments a Long Short-Term Memory (LSTM) forecaster with a causal, time-aligned anomaly indicator generated by a Hyperdimensional Computing (HDC) classifier. The HDC module encodes sliding windows of the load series into highdimensional representations and performs prototype-based associative inference. During training, the LSTM uses the available anomaly labels as an additional input feature, whereas during inference these labels are replaced by the causal indicators predicted by the HDC classifier. Experiments on five years of hourly Brazilian electricity-consumption data with controlled anomaly injections show consistent improvements over an otherwise identical LSTM baseline without anomaly features. The proposed configuration reduces RMSE by 14.9% over the full test horizon and by 18.5% on peak-demand intervals. These results indicate that conditioning the forecaster on lightweight HDC-derived anomaly context improves robustness against irregular load behavior with minimal additional model complexity.

Paper Nr:	64
Title:	Comparing Bandwidth Allocation of Deep Reinforcement Learning and OFDMA in Wireless Networks: A Case of Wi-Fi 7 (IEEE 802.11be)
Authors:	Daniel Kiguru Ruiru, Nicolas Jouandeau and Dickson Owuor
Abstract:	This paper compares Deep Reinforcement Learning (DRL) agents with OFDMA in allocating bandwidth. OFDMA as used in Wi-Fi 7 (802.11be) is presented in this paper where its primarily characterized by flexible Resource Units (RUs) that can be allocated to different devices. To simulate realistic, realworld scenarios, dynamic and volatile user conditions are incorporated into the experiments. In the results, a Transformer based Deep Q-Network outperforms all other algorithms. This algorithm gets the best adaptability score (92%) and volatility handling capability score (89%). These values were the ultimate comparison methods following a series of evaluations that tested all the algorithms on metrics of throughput, fairness, delay, energy efficiency, and satisfaction. This paper concludes that DRL algorithms are ideal for dynamic and volatile environments while OFDMA techniques are good for static and predictable conditions.

Paper Nr:	66
Title:	Spike-Based Belief Propagation in Nonlinear Dynamical Systems
Authors:	Sepideh Adamiat, Hongye Wang, Wouter M Kouw and Bert de Vries
Abstract:	This paper presents a Bayesian control framework that integrates spikebased dynamics with probabilistic inference for adaptive control. Bayesian inference is widely regarded as a core computational principle of brain function, providing a normative framework for perception, decision-making, and learning under uncertainty. By combining a biologically inspired spiking neural model with Bayesian inference principles, we propose a brain-like control algorithm capable of operating in uncertain environments. The architecture is hybrid, embedding a spiking neural realization of the nonlinear state-transition factor within a numerical message-passing framework. We use the mountain car parking problem as a benchmark with non-linear dynamics. Our results demonstrate that the proposed controller can successfully update states online, at each control step, and generate goal-directed action plans through spike-driven dynamics. The results highlight the proposed model’s potential as a bridge between computational neuroscience and probabilistic control theory.

Paper Nr:	81
Title:	Brain Tumor Segmentation in T1-Weighted MRI Using Improved Quantum Particle Swarm Optimization
Authors:	Divya Sahgal, Roman Senkerik and Zuzana Kominkova Oplatkova
Abstract:	Accurate tumor delineation in neuroimaging remains a challenge for both skilled radiologists and automated systems, especially when dealing with the morphological diversity of intracranial masses. The possibility of using adaptive parameter selection based on unique scan characteristics to address this variability was investigated using Improved Quantum Particle Swarm Optimization (IQPSO). The proposed pipeline combines dual-channel feature extraction covering intensity hyperintensities and gradient discontinuities, bias field correction calibrated for T1-weighted sequences, and quantum-inspired search that strikes a balance between exploration and exploitation over 30 iterations. 94.0% classification accuracy was shown in validation on 50 consecutive cases from single institution, including 33 age-matched controls and 17 pathology-confirmed tumors. It's interesting to note that while there were no false positives among normal scans (specificity 100%), 14 out of 17 tumors were detected (sensitivity 82.4%). The average per-case runtime on standard CPU hardware was 5.44 seconds, and the mean dice overlap reached 0.160 for positive results. These results establish IQPSO as a useful screening supplement in settings with limited resources where deep learning deployment is still not feasible.

Paper Nr:	85
Title:	A Reliability Evaluation Methodology for Quantized Vision-Language Model Deployment
Authors:	Rupert Urbanski and Ralf Peters
Abstract:	On-premises and edge deployment of vision-language models often relies on post-training quantization, yet configuration choices typically rest on aggregate accuracy, which masks its heterogeneous reliability effects. This paper develops a reliability evaluation methodology, grounded in design science and organized around Wang and Strong’s four data quality categories, that brings a structure to a problem the ML literature has approached metric by metric. The methodology is instantiated on Qwen2.5-VL across four scales, ten weight configurations at four- and eight-bit and three component-protection conditions. Eight-bit is essentially lossless on accuracy. At four-bit especially, the methodology identifies accuracy-equivalent configurations that diverge on flip rate, calibration and logit distribution. The category structure holds under construct-validity testing at every scale, and the vision-encoder protection heuristic is recast as a decision conditioned on format, component and scale.

Paper Nr:	92
Title:	Extended Hardware-Aware Neural Architecture Search Space for Transformer Networks Using Proxy Evaluations
Authors:	Steven Colleman, Rafaella Octaviani de Assuncao and Toon Goedemé
Abstract:	Hardware-Aware Neural Architecture Search (HW-NAS) has become an important approach for designing neural networks that jointly optimize predictive performance and deployment efficiency. While existing HW-NAS methods have been extensively explored for convolutional neural networks, relatively little attention has been given to transformer-based architectures despite their growing dominance in computer vision. This work extends the NSGA-Net framework with a collection of hardware-efficient Vision Transformer primitives, enabling the evolutionary discovery of hybrid CNN--Transformer architectures. To address the resulting increase in search space complexity, we introduce a proxy-filtered evolutionary search strategy that combines lightweight proxy metrics with high-fidelity hardware evaluations. The proposed approach is evaluated on the EuroSAT and Oxford Flowers102 image classification benchmarks across three target hardware platforms: Analog In-Memory Computing (IMC), Digital IMC, and TPU accelerators. In total, twelve multi-objective NAS experiments were conducted, jointly optimizing classification error and either energy consumption or inference latency. The results demonstrate that the proxy-filtered search substantially improves exploration efficiency while maintaining a fixed evaluation budget. Furthermore, the experiments reveal strong interactions between task complexity, hardware characteristics, and architectural topology. In particular, lightweight attention mechanisms emerged as highly effective alternatives to conventional convolutions for fine-grained visual recognition under strict hardware constraints. These findings highlight the potential of hardware-aware evolutionary search for discovering deployable transformer-based architectures tailored to diverse accelerator platforms.

Paper Nr:	93
Title:	Tool Design for Analysis of Flexible Depth-First Schedulings of Residual-Based Convolutional Neural Networks
Authors:	Kyryll Filippenkov, Steven Colleman and Toon Goedemé
Abstract:	Depth-first scheduling makes dedicated CNN hardware stream high-resolution images line by line, drastically reducing on-chip memory and off-chip I/O data transfers relative to layer-by-layer execution. These gains, however, have only been established for straightforward, sequential network topologies. UNets break that assumption: their residual connections force feature data to remain on-chip far longer than a single kernel window, and existing closed-form analyses of the resulting overhead are valid only when the residual branch contains no convolutional layer, which rules many realistic UNet blocks. As a novelty, we present a tool that allows to investigate this restriction. It models a UNet subnetwork as a set of branches that share one input feature map and merge element-wise, where any branch (including the skip path) may itself chain convolution and up/down-sampling layers. For such a subnetwork, the tool explores the space of valid line-based depth-first schedulings and returns the exact Pareto frontier of the trade-off between peak on-chip memory and total I/O communication, reporting for every operating point a concrete, replayable schedule. This turns the single closed-form line count of prior work into a full design-space curve, letting hardware designers select the memory/bandwidth operating point that best matches their accelerator and measure how strongly residual connections erode the memory and I/O savings that depth-first scheduling delivers for plain, sequential CNNs.

Paper Nr:	96
Title:	Security and Privacy in Large Language and Foundation Models: A Survey on Defensive Strategies for GenAI Models
Authors:	Giorgio Piccardo, Giuseppe F. Italiano and Alessio Martino
Abstract:	Large Language Models (LLMs) and Foundation Models (FMs) have revolutionized natural language processing, enabling remarkable advancements in language understanding and generation. However, their deployment also raises critical concerns around data privacy, adversarial attacks, and the potential for misuse. This survey explores the defensive side of the security and privacy landscape surrounding LLMs and FMs, with a particular focus on the perspective and guidelines provided by the National Institute of Standards and Technology (NIST). Following the PRISMA 2020 review protocol, we examine the defence strategies and mitigation techniques proposed to enhance the security and privacy of these models, organized around the NIST taxonomy of threats: availability, integrity, privacy, and abuse attacks. As a complement to recent surveys on GenAI attacks, this work consolidates the state of the art of mitigation strategies and, through an explicit attack–defence crosswalk, maps each defence family onto the threat it counters. We summarise the evidence in a comparative table and a taxonomy, and critically examine robustness and limitations, foremost the recurring finding that many defences are weakened by adaptive attacks. Beyond text-only LLMs, we cover defences specific to vision–language and text-to-image FMs, and we discuss supply-chain and deployment-level mitigations within the NIST framework. We close by outlining open challenges and research directions toward more secure and privacy-preserving generative AI systems, together with practical implications for deployment.

Paper Nr:	99
Title:	Video–Subtitle Automatic Alignment for Sign Language Datasets: Challenges Illustrated on Matignon-LSF Dataset
Authors:	Julie Halbout, Diandra Fabre, Michèle Gouiffès and Annelies Braffort
Abstract:	Despite its critical importance, automatic video-text alignment in sign language (SL) research remains largely unexplored, with few machine learning frameworks proposed to date. Addressing this gap is essential to advance Sign Language-to-Text translation, a task currently hindered by the scarcity of high-quality aligned datasets. This paper investigates these alignment challenges using the Matignon-LSF dataset, a multimodal corpus comprising spoken French discourse, written French subtitles, and French Sign Language (LSF) interpretations. Three distinct alignment methodologies are evaluated, including the fine-tuning of state-of-the-art models pre-trained on large-scale SL datasets, to enhance data availability and alignment precision. The proposed methods are quantitatively evaluated, and their performance is further analyzed through a detailed qualitative discussion. Finally, we analyze alignment discrepancies to elucidate the complex phenomena that occur during interpretation, underscoring the need for specialized efforts in alignment.

Paper Nr:	121
Title:	Stage-Aware Evaluation of Federated Learning with Individualized Updates in Non-IID Environments
Authors:	Mortesa Hussaini, Jan Theiß and Anthony Stein
Abstract:	In Federated Learning (FL) with heterogeneous data distributions, local models tend to converge toward client-specific optima, deviating from the global data distribution. Aggregation methods, such as FedAvg, often suffer from client drift, producing a global model that is suboptimal for most clients. Personalized Federated Learning (PFL) addresses local performance, yet out-of-distribution generalization, a key benefit of global aggregation and a critical component of robustness, remains insufficiently captured in existing evaluation frameworks. In practice, such FL systems operate as decentralized ensembles of clients in environments that are not only statically heterogeneous, but also subject to dynamic shifts in data distributions and client participation, conditions that necessitate adaptive behavior at the level of individual clients. This work introduces a stage-aware evaluation scheme that distinguishes three semantically distinct evaluation points within a single communication round: the post-aggregation global model (G), the individualized pre-training update (L1) and the post-local-fine-tuning model state (L2). Critically, stage L1 is frequently overlooked in existing PFL literature. Complementing this evaluation framework, we propose Federated Learning with Individualized Updates (FLIU), a lightweight extension of FedAvg that introduces a client-specific personalization step governed by an adaptive, data-driven personalization factor. Among other things, our empirical evaluation reveals that (i) the proposed harmonic metric AccH exposes performance imbalances invisible to local or global accuracy alone, (ii) stage L1 already achieves a competitive local–global balance before any local training and (iii) additional local training at stage L2 can substantially erode out-of-distribution generalization.

Paper Nr:	34
Title:	Transformer-Based Deep Learning for Stroke Outcome Prediction: Addressing Severe Class Imbalance
Authors:	María N. Moreno-García, Alberto San Segundo Gutiérrez, Araceli Rodríguez Vico and Fernando Sánchez-Hernández
Abstract:	Stroke outcome prediction from clinical tabular data remains a challenging task due to heterogeneous patient variables, limited dataset sizes, and the severe class imbalance commonly present in medical outcome distributions. Conventional machine learning approaches often achieve high overall accuracy while showing poor sensitivity toward minority classes, which frequently correspond to clinically critical patient conditions. Although deep learning methods have shown strong representation learning capabilities, their application to clinical tabular data is often limited by small dataset sizes, overfitting risks, reduced interpretability, and difficulties in handling highly imbalanced data. To address these limitations, this work proposes a transformer-based deep learning framework for stroke outcome prediction. The proposed methodology leverages self-attention mechanisms to capture contextual relationships among heterogeneous clinical variables while integrating imbalance-aware learning strategies to improve minority class recognition. By combining transformer-based contextual representation learning with specialized optimization techniques for imbalanced data, the framework aims to achieve a more balanced predictive performance across outcome categories. The proposed approach was evaluated using stratified cross-validation and compared against multiple state-of-the-art methods, including conventional machine learning algorithms, resampling techniques, attention-based neural models, and GAN-based balancing strategies. Experimental results show that the proposed architecture substantially improves the classification performance of minority outcome classes while maintaining competitive performance for majority classes. These findings highlight the potential of transformer-based architectures combined with imbalance-aware learning strategies for robust clinical tabular prediction and stroke prognosis assessment.

Paper Nr:	53
Title:	Severity-Aware Sliding Window Classification with Overlap for DDoS Detection
Authors:	Eugenio Borrini, Enrico De Santis and Antonello Rizzi
Abstract:	Distributed Denial-of-Service (DDoS) attacks represent a major class of availability threats in modern networks, aiming to deny access to web resources to authorized users by exhausting computational or bandwidth resources. Due to these intrinsic DDoS characteristics, classifying the individual packets is often extremely difficult and, in some cases, ineffective. As a solution, a framework based on classifying a fixed-width sliding window has been developed. Windows have been labeled using different metrics based on the percentage of attack packets inside the sliding window. The main advantage is that a threat is detected only when it exceeds a user-defined threshold - which can be considered as network tolerance - avoiding non-harmful attacks and preventing instability during training caused by benign-attack interleaved flows.

Paper Nr:	58
Title:	Perceptually Accurate Prediction of Nonlinear Beverage Color Mixing in CIELAB Space Using Neural Networks
Authors:	Chiheb Naffeti, Elyes Maherzi and Yassine Hachaichi
Abstract:	Accurate prediction of color outcomes in beverage mixtures remains challenging due to nonlinear interactions between pigments, pH-dependent chromatic transformations, and light scattering, which limit the accuracy of conventional linear interpolation and physics-based mixing models. This paper presents a data-driven framework that predicts the CIELAB coordinates (L∗, a∗, b∗) of binary beverage mixtures from the volumetric mixing proportion and the CIELAB coordinates of the constituent beverages, using Multi-Layer Perceptron (MLP) regression models trained and validated on a large-scale dataset of 103,011 beverage mixing observations. Two optimization strategies, L-BFGS and Adam, were compared under identical network architectures. The L-BFGS-optimized model achieved the best predictive performance, with a test coefficient of determination R 2 = 0.9797 and, based on the CIEDE2000 perceptual color difference metric, a mean prediction error of ∆E00 = 0.998 and a 90th-percentile error of 2.441. These results indicate that the great majority of predictions fall within industrially acceptable color-matching tolerances, and that the proposed lightweight architecture offers a practical alternative to traditional analytical color prediction methods for realtime beverage formulation and quality control.

Paper Nr:	71
Title:	Game-Theoretic Fair Allocation in MPC-Based Management of RECs
Authors:	Danial Zendehdel, Enrico De Santis and Antonello Rizzi
Abstract:	Renewable Energy Communities (RECs) enable prosumers to exchange locally generated renewable energy, but designing fair and efficient coordination mechanisms remains challenging. This work proposes a three-stage decentralized framework for peer-to-peer (P2P) energy trading in RECs with battery storage. In Stage~1, each household solves a local Model Predictive Control (MPC) problem to determine battery-aware P2P bids. In Stage~2, a market coordinator clears trades using either proportional allocation or disagreement-cost weights, the latter motivated by the asymmetric Nash bargaining solution. In Stage~3, agents finalize battery schedules given the allocated P2P commitments. Unlike centralized approaches, the framework decomposes the community problem into per-agent subproblems and requires agents to disclose only aggregate bid quantities rather than full load, generation and storage profiles. We argue that fairness in RECs should be defined primarily through \textit{equity}, rather than mere equality of distribution. We define efficiency as minimizing total energy waste, demonstrating that high P2P trading volume does not guarantee efficiency. Comparative experiments using real household data show that disagreement-cost weighting matches the efficiency of centralized Fair-MPC ($47.9\%$ vs.\ $47.1\%$) while improving equity by $50\%$, despite trading $75\%$ less P2P volume. This indicates that decentralized cooperative mechanisms can match centralized optimization on efficiency while substantially improving need-based fairness.

Paper Nr:	90
Title:	Self-Organized Learning in Oscillatory Neural Networks with Memristive Signed Couplings
Authors:	Riley Acker, Aman Desai, Garrett Kenyon and Frank Barrows
Abstract:	Oscillatory neural networks (ONNs) have emerged as a promising neuromorphic architecture, leveraging dynamical systems to perform computation and represent information through phase relationships. Their interactions can be designed to support intrinsic energy-minimizing dynamics, enabling tasks such as associative memory and optimization, and positioning them as a candidate architecture for continuous learning and inference. We present a neuromorphic primitive implemented using memristive inhibitory couplings for autonomous learning in oscillatory neural networks. We validate the proposed architecture through circuit simulations, demonstrating its ability to perform denoising in an auto-associative memory task. While Hopfield and Ising models routinely assume signed weights, integrating signed couplings with continuous, device-driven learning remains a circuit-design challenge. The novelty of this work lies in a circuit mechanism for learning signed effective couplings through local memristive dynamics. A fixed excitatory pathway competes with a volatile memristive inhibitory pathway, allowing local, phase-dependent memristive dynamics to modify both coupling magnitude and sign. A physically implementable route to inhibitory (negative) weights is particularly valuable: it expands the class of attractor structures accessible to oscillator networks beyond purely synchronous couplings, and supports phase-coded memories where anti-phase constraints are not merely transiently enforced during training but can persist autonomously after release. Circuit simulations demonstrate learning and autonomous recall of in-phase and anti-phase relationships through local memristive adaptation, while theoretical analysis establishes the necessity of negative effective couplings for anti-phase stability within the considered phase-reduced model.

Paper Nr:	94
Title:	Motif Discovery in Time Series: A Survey
Authors:	Enrico De Santis, Alessio Martino, Vanessa Pennazzi and Antonello Rizzi
Abstract:	Motif discovery is a core task in time series data mining, aimed at identifying repeated, conserved, or otherwise informative subsequences in sequential data. This paper provides a compact survey of time series motif discovery methods, with emphasis on algorithmic principles, scalability, software availability, and recent methodological trends. After introducing the minimum background needed to define time series, subsequences, similarity measures, representations, and motifs, we organize the literature along a chronological and methodological axis. Early approximate and symbolic approaches are discussed as historical milestones, while particular attention is devoted to exact, scalable, matrix-profile-based, multivariate, variable-length, and neural representation-based methods. The survey also summarizes available implementations and representative application domains, including biological sequences, motion data, energy monitoring, automotive signals, and music information retrieval. The resulting synthesis is intended to support researchers and practitioners in selecting motif discovery techniques according to data size, dimensionality, required accuracy, interpretability, and computational constraints.

Paper Nr:	107
Title:	Analytical Framework Modelling Stereo Depth Estimation Disparity-Sweep Schedulings Using Depth-First
Authors:	Steven Colleman, Xin Huang and Toon Goedemé
Abstract:	Real-time depth estimation using stereo-based Convolutional Neural Networks (CNNs) is a critical component in autonomous driving and robotic vision. However, scaling these networks to high-resolution stereo pairs with large disparity ranges (Dmax) introduces severe bottlenecks in terms of computational latency, on-chip SRAM footprint, and off-chip DRAM I/O traffic. While prior work successfully optimized hardware mapping for single-disparity steps, practical systems require simultaneous evaluation of multiple disparities, which inherently references highly redundant image regions. This paper presents a systematic framework analyzing twelve distinct disparity-sweep scheduling strategies, categorized along two orthogonal dimensions: disparity scheduling (perdisparity, full batch, or tiled) and precomputation levels. Each scheduling strategy makes use of depth-first. We derive rigorous algebraic formulations for each strategy across computational latency, external I/O communication cost, and peak on-chip SRAM capacity. Our analysis reveals a highly complex, multi-dimensional design space where scheduling choices dictate strict Pareto-optimal trade-offs. Specifically, we demonstrate that full-batch sliding window strategies eliminate redundant DRAM inputs but heavily penalize post-merge SRAM caching, whereas parameterized tiling provides a critical runtime tuning knob to compress memory footprints at a proportional latency penalty. Ultimately, this comprehensive framework allows hardware designers to select optimal scheduling strategies mapped precisely to SRAM-constrained, energy-constrained, or throughput-optimized hardware regimes.

Paper Nr:	123
Title:	Unified Audio-Visual Predictive Coding: A Generative Framework for Convergent Machine Listening
Authors:	Chetas Parekh and Sanchita Ghose
Abstract:	Navigating a complex auditory world requires more than disjoint feats of signal processing: it demands a unified, embodied perception of what sounds are and where they originate. The fields of Sound Source Localization (SSL) and Sound Source Separation (SSS) have achieved remarkable progress under deep learning, yet they remain siloed disciplines, optimized against disjoint benchmarks. We argue that this separation is an artifact of signal-processing engineering history rather than a property of intelligent perception, and that it is fundamentally incompatible with embodied, agentic artificial intelligence systems, which must continuously predict and map sensory streams onto physical space in order to act. We advance Audio-Visual Predictive Coding (AVPC) as a unifying computational framework: a self-supervised generative system in which cross-modal predictive error between audio and visual streams serves as an intrinsic loss function, and in which “separation” and “localization” emerge as joint byproducts of a single predictive objective rather than as independently engineered targets. Surveying developments between 2023 and mid-2026, we show that state-of-the-art architectures are already converging implicitly toward this formulation, and we situate the proposal within the broader discourse on generative and agentic AI. We further examine the resulting design space through the lens of Explainable AI (XAI), characterizing the trade-off between the representational capacity of deep generative architectures and the interpretability and safety guarantees offered by unimodal, physics-based methods. We conclude by urging the computational intelligence community to formalize this convergence, replacing task-separated benchmarks with holistic evaluation protocols for embodied, multimodal agents.