Image aesthetic attribute assessment provides explainable outputs for Image Aesthetic Quality Assessment (IAQA), evaluating attributes such as rule of thirds, symmetry, and lighting. These attributes are context-dependent, as their importance varies across different photography scenes. However, most attribute-based IAQA methods remain scene-agnostic, limiting their ability to model scene–attribute dependencies. We propose a scene-aware IAQA model based on a vision transformer that extracts multi-level features and integrates learned scene embeddings within a two-tower module to capture both general and scene-specific patterns, with adaptive gating for context-aware fusion. Experimental results show improved correlation for both overall score and attribute prediction, outperforming state-of-the-art attribute-based IAQA methods.

Panoramic X-ray images are essential for dental caries detection, yet they inherently suffer from geometric distortions that complicate accurate early diagnosis. To address this, we propose a Geometry-Biased transformer that explicitly models spherical geometry. Our approach integrates equirectangular relative position embedding, distance-based attention scoring, and equirectangular-aware attention rearrangement to significantly enhance spatial feature representation. This enables the model to effectively capture both local caries lesions and global dental arch structures despite distortions. We rigorously evaluated our model primarily on a hospital-scale dataset, achieving a high accuracy of 94.90% and an AUC of 0.9603. Furthermore, extensive comparative analyses across diverse publicly available dental image datasets demonstrate the superior generalization and competitive performance of our method. Our findings highlight that geometry-aware transformers offer a robust and automated tool, revolutionizing high-precision dental caries detection and potentially other medical imaging diagnostics.

Depression detection benefits from combining neurological and behavioral indicators, yet integrating heterogeneous modalities such as EEG and interview audio remains challenging. We propose a transformer-based multimodal framework that jointly models spectral, spatial, and temporal EEG features alongside linguistic and paralinguistic cues from interviews. By employing synchronized multi-head cross-attention and self-attention mechanisms, the model effectively captures intra- and inter-modal correlations. In addition, a flexible temporal sequence matching strategy reduces EEG channel requirements, enhancing device portability. Evaluated on the MODMA and DAIC-WOZ datasets, our approach achieves superior performance compared to state-of-the-art models, with a 4.7% improvement in accuracy and a 10% increase in precision. These results demonstrate the potential of the proposed framework for accurate, scalable, and cost-effective depression detection in both clinical and real-world settings.

Computing-in-Memory (CIM) architectures using memristive crossbar arrays enable energy-efficient AI acceleration. Analog non-idealities, such as IR drop and nonlinearity, impose design constraints that existing simulators cannot capture and thus explore effectively. Current approaches sacrifice either modeling accuracy or simulation speed, preventing systematic design space exploration. In this paper we propose X-Sim, a crossbar simulator that resolves this trade-off through a modular architecture. Our approach decouples device physics from circuit analysis using a fixed-point scheme, avoiding expensive Jacobian computations while preserving device fidelity. X-Sim delivers SPICE-level accuracy (< 1% error) with up to 200× speedup over physics-based simulators. This enables quick and systematic design space exploration across thousands of configurations, guiding reliable system design. X-Sim will be released as open source.

Image Aesthetic Quality Assessment (IAQA) spans applications such as the fashion industry, AI-generated content, product design, and e-commerce. Recent deep learning advancements have been employed to evaluate image aesthetic quality. A few surveys have been conducted on IAQA models; however, details of recent deep learning models and challenges have not been fully mentioned. This article aims to fill these gaps by providing a review of deep learning IAQA over the past decade, based on input, process, and output phases. Methodologies for deep learning-based IAQA can be categorized into general and task-specific approaches, depending on the type and diversity of input images. The processing phase involves considerations related to network architecture, learning structures, and feature extraction methods. The output phase generates results such as scoring, distribution, attributes, and description. Despite achieving a maximum accuracy of 91.5%, further improvements in deep learning models are still required. Our study highlights several challenges, including adapting models for task-specific methodology, accounting for environmental factors influencing aesthetics, the lack of substantial datasets with appropriate labels, imbalanced data, preserving image aspect ratio and integrity in network architecture design, and the need for explainable AI to understand the causative factors behind aesthetic judgments.

Accurate psychological text analysis on embedded devices is essential for privacy-preserving mental health monitoring; however, deploying transformer-based models in such constrained environments remains challenging due to limited memory, computational power, and energy budgets. We present a multimodal framework that integrates a lightweight Transformer (DistilBERT) with a CNN-based image branch (ResNet-50) accelerated by the Hailo-8L edge AI board. Text inputs are processed both as sequences and as two-dimensional image-like representations, enabling robust handling of noisy, informal text containing emojis and misspellings. Evaluations on the Depression Twitter dataset and the SST-2 benchmark demonstrate that the fusion model consistently outperforms text-only and image-only baselines, achieving up to 91.6% accuracy while remaining feasible for deployment on a Raspberry Pi 5. Overall, the proposed design offers a practical balance between efficiency and accuracy, enabling real-world, privacy-aware psychological monitoring directly on personal embedded devices.

Accurate and secure classifying informal documents related to mental disorders is challenging due to factors such as informal language, noisy data, cultural differences, personal information and mixed emotions. Conventional deep learning models often struggle to capture patterns in informal text, as they miss long-range dependencies, explain words and phrases literally, and have difficulty processing non-standard inputs like emojis. To address these limitations, we expand data dimensionality, transforming and fusing textual data and signs from a 1D to a 2D space. This enables the use of pre-trained 2D CNN models, such as AlexNet, Restnet-50, and VGG-16 removing the need to design and train new models from scratch. We apply this approach to a dataset of social media posts to classify informal documents as either related to suicide or non-suicide content. Our results demonstrate high classification accuracy, exceeding 99%. In addition, our 2D visual data representation conceals individual private information and helps explainability.

Multimodal AI systems increasingly rely on biomedical signals such as EEG and eye movement data for emotion recognition. However, these models face challenges including limited training data, inter-subject variability, session-specific spurious correlations, and incomplete modality representation, all of which reduce generalization and reliability. We propose TEREE, a multimodal transformer-based model that integrates temporal, spatial, and spectral EEG features with eye movement data. To mitigate session-specific artifacts, Bayesian Spurious Correlation Minimization (BSCM) is applied. In addition, a holistic multimodal processing strategy enables robust handling of incomplete data. The model was trained and evaluated using the SEED and SEED-FRA benchmark datasets under one-to-one and multi-to-one transfer paradigms. TEREE achieved state-of-the-art performance, with average multi-to-one transfer accuracies of 97.7% on SEED and 98.8% on SEED-FRA. Ablation studies confirmed that fusing EEG with eye movement features consistently improved accuracy compared to unimodal baselines. Standard deviations across repeated experiments were below 5%, indicating stability. By addressing inter-subject variability, spurious correlations, and incomplete modality issues, TEREE enhances the robustness and generalization of emotion recognition systems. These findings suggest that multimodal transformer-based models can substantially improve the reliability of affective computing applications such as human–computer interaction and mental health monitoring.

Full-System (FS) simulation is essential for performance evaluation of complete systems that execute complex applications on a complete software stack consisting of an operating system and user applications. Nevertheless, they require careful fine-tuning against real hardware to obtain reliable performance statistics, which can become tedious, error-prone, and time-consuming with typical trial-and-error approaches. We propose a novel, streamlined, component-level calibration methodology to address these shortcomings to validate FS simulation models. Our methodology greatly accelerates the validation process without sacrificing accuracy. It is Instruction Set Architecture (ISA)-agnostic, and can tackle hardware specifications at different levels of detail. We demonstrate its effectiveness by validating FS models against both open-hardware and IP-protected (closed hardware) RISC-V silicon, achieving a mean error of 19%-23% for the SPEC CPU2017 suite in the two cases. We introduce the first open-source RISC-V-based FS-validated simulation models with a complete and replicable methodology.

Processing graphs on a large scale presents a range of difficulties, including irregular memory access patterns, device memory limitations, and the need for effective partitioning in distributed systems, all of which can lead to performance problems on traditional architectures such as CPUs and GPUs. To address these challenges, recent research emphasizes the use of Field-Programmable Gate Arrays (FPGAs) within distributed frameworks, harnessing the power of FPGAs in a distributed environment for accelerated graph processing. This paper examines the effectiveness of a multi-FPGA distributed architecture in combination with a partitioning system to improve data locality and reduce inter-partition communication. Utilizing Hadoop at a higher level, the framework maps the graph to the hardware, efficiently distributing pre-processed data to FPGAs. The FPGA processing engine, integrated into a cluster framework, optimizes data transfers, using offline partitioning for large-scale graph distribution. A first evaluation of the framework is based on the popular PageRank algorithm, which assigns a value to each node in a graph based on its importance. In the realm of large-scale graphs, the single FPGA solution outperformed the GPU solution that were restricted by memory capacity and surpassing CPU speedup by 26x compared to 12x. Moreover, when a single FPGA device was limited due to the size of the graph, our performance model showed that a distributed system with multiple FPGAs could increase performance by around 12x. This highlights the effectiveness of our solution for handling large datasets that surpass on-chip memory restrictions.

Memristor technology has shown great promise for energy-efficient computing [1] , though it is still facing many challenges [1 , 2]. For instance, the required additional costly electroforming to establish conductive pathways is seen as a significant drawback as it contributes to power and area overheads, and limited device endurance. In this work, we propose a novel forming-free HfO₂ -based ReRAM device with low operating voltages , multi-level capability , and less sensitivity to device-to-device (D2D) and cycle-to-cycle (C2C) variations. The device is fabricated using CMOS-compatible processes, excluding the undesirable complex steps mandatory to manufacture the state-of-the-art forming-free devices [3, 4, 5]. This is accomplished by utilizing the desirable formation energy of Pd-O bonds [6, 7], which creates conducting paths at room temperature while maintaining the analog switching ability of the devices. The proposed ReRAM device holds a great value for dense memories and energy-efficient compute architectures.

Neural Networks (NN) are often trained offline on large datasets and deployed on specialised hardware for inference, with a strict separation between training and inference. However, in many realistic applications the training environment differs from the real world, or data arrives in a streaming fashion and is continuously changing. In these scenarios, the ability to continuously train and update NN models is desirable. Continual learning (CL) algorithms allow training of models on a stream of data. CL algorithms are often designed to work in constrained settings, such as limited memory and computational power, or limitations on the ability to store past data (e.g, due to privacy concerns or memory requirements). High-energy physics experiments are developing intelligent detectors, with algorithms running on computer systems located close to the detector to meet the challenges of increased data rates and occupancies. The use of NN algorithms in this context is limited by changing detector conditions, such as degradation over time or failure of an input signal which might cause the NNs to lose accuracy leading, in the worst case to the loss of interesting events. CL has the potential to solve this issue, using large amounts of continuously streaming data to allow the network to recognise changes, and to learn and adapt to detector conditions. It has the potential to outperform traditional NN training techniques as not all possible scenarios can be predicted and modelled in static training data samples. However, NN training is computationally expensive and when combined with the strict timing requirements of embedded processors deployed close to the detector, current state-of-the-art offline approaches cannot be directly applied to the real-time systems. Alternatives to typical backpropagation-based training that can be deployed on FPGAs for real-time data processing are presented, and their computational and accuracy characteristics are discussed in the context of High-Luminosity LHC.

Collision recovery is considered one of the main potentials for improving bulk reading speed in the UHF radio identification (RFID) system. The collision occurs when two or more tags reply at the same time. State-of-the-art collision recovery algorithms rely on perfect channel state information (CSI), which is unrealistic. Moreover, these algorithms neglect the rate tolerance, that is, the uncertainty in the bit rate of the tag responses. This paper presents a novel algorithm to estimate all tag parameters, such as tag data rate, number of tags per slot, and CSI, using the Matrix Pencil Method (MPM). The proposed system is compatible with the existing EPCglobal Class-1 Gen-2 RFID standard. The accuracy of the proposed algorithm is evaluated with the different encoding schemes of the tag. The estimated parameters are forwarded to MIMO linear receivers to recover the colliding slots to speed up the identification process. The performance of the three linear receivers; Zero Forcing (ZF), Minimum Mean Square Error (MMSE), and Maximum Likelihood (ML) receivers, is evaluated when the CSI is estimated using the proposed algorithm. All receivers perform similarly due to their different sensitivity to the CSI estimation error. The proposed collision recovery system reduces the time for reading a bulk by at least 13% compared to the literature.

Applications of Binary Neural Networks (BNNs) are promising for embedded systems with hard constraints on energy and computing power. Contrary to conventional neural networks using floating-point datatypes, BNNs use binarized weights and activations to reduce memory and computation requirements. Memristors, emerging non-volatile memory devices, show great potential as a target implementation platform for BNNs by integrating storage and compute units. However, the efficiency of this hardware highly depends on how the network is mapped and executed on these devices. In this paper, we propose an efficient implementation of XNOR-based BNN to maximize parallelization. In this implementation, costly analog-to-digital converters are replaced with sense amplifiers with custom reference(s) to generate activation values. Besides, a novel mapping is introduced to minimize the overhead of data communication between convolution layers mapped to different memristor crossbars. This comes with extensive analytical and simulation-based analysis to evaluate the implication of different design choices considering the accuracy of the network. The results show that our approach achieves up to 5× energy-saving and 100× improvement in latency compared to baselines.