Multimodal learning systems typically assume strong alignment between data modalities, an assumption frequently violated in real-world scenarios such as ecological monitoring, where visual and acoustic observations are weakly or inconsistently paired. This paper proposes a hybrid quantum-classical fusion framework that integrates independently learned modality-specific representations through a Parameterized Quantum Circuit (PQC). The proposed Quantum Fusion Layer (QFL) is designed to capture high-order cross-modal interactions under weak alignment conditions. Experimental evaluation on a multimodal bird species classification task shows that the proposed framework outperforms classical fusion baselines, achieving Accuracy and Macro F1-score of 0.9875 on a held-out test set. Additional analyses comprising stratified cross-validation, ablation studies, robustness evaluation under controlled noise conditions, a parameter-matched classical MLP control, and cross-domain generalization on a geographically out-of-distribution benchmark (Bird-SEA10) of 238 multimodal pairs from Singapore and Thailand indicate stable performance across data partitions, resilience to moderate input degradation, and meaningful generalization under geographic domain shift (cross-domain Accuracy =0.8445, Macro F1 =0.8382, ΔF1=−0.1493 relative to in-domain performance). The parameter-matched MLP control achieves Accuracy of 0.2125 and Macro F1 of 0.0897, suggesting that the observed performance advantage originates from the structural properties of the quantum circuit rather than from reduced parameter capacity. These findings indicate that quantum-enhanced feature transformations may provide an effective mechanism for multimodal fusion under weak alignment, with potential applicability to broader multimodal learning problems involving heterogeneous and imperfectly aligned data sources.

The increasing complexity and dynamic nature of Industrial Internet of Things (IIoT) demand scalable, adaptive, intuitive, and real-time automation frameworks. This paper presents a systematic literature review (SLR) of edge- and cloud-based collaborative learning frameworks for predictive maintenance and smart manufacturing tasks. In this SLR, we highlight the under-utilization of distributed computational architectures that provide complete automation support (design and run-time), flexibility, scalability, and inter- & intra-cloud service exchange while adhering to security management and integrity principles for solving IIoT tasks using modern artificial intelligence (AI) models at the edge/cloud. Recently, many IIoT applications have been designed using AI models that require robust, low-latency, and data-secure frameworks. This demand drives a trend toward distributed computational architectures in which data storage and processing are partially or fully decentralized. Common paradigms addressing this resource distribution include edge computing, federated learning, and private or hybrid clouds. We analyze 50 recent studies against IoT characteristics, AI performance metrics, and network/system management requirements. Our findings reveal underutilization of distributed architectures that support automation, interoperability, and security. While most solutions rely on centralized or hybrid clouds, fewer than 5% adopt federated or transfer learning, and over 60% remain dependent on supervised models. We also introduce a comparative perspective on network and security management, showing that local/private cloud implementations can reduce control-plane overhead and synchronization latency, though gaps persist in dynamic bandwidth allocation and zero-trust adoption. Finally, we benchmark our previously proposed local cloud-based collaborative learning (CCL) model against state-of-the-art solutions, highlighting its strengths in automation and interoperability, as well as limitations in adaptive computation and intelligent offloading. This review identifies the research gaps and opportunities for integrating collaborative AI, secure automation, and hybrid architectures to meet Industry 5.0 objectives of resilience, sustainability, and human-centricity.

Aside from common challenges in aerial object detection, i.e., small, sparse targets and computational power limitations, detecting vehicles from UAV images in the Nordic regions faces strong visibility challenges and domain gaps caused by diverse levels of snow coverage. Although data annotation is expensive, unannotated data is cheaper to collect by simply flying the drones. Hence, we propose a Sideload-Contrastive-Learning-Adaptation (SCLA) framework to improve lightweight detection model performance utilizing cheap unannotated data. First, we propose the Feature-Map-Patch Contrastive Learning (FM-PaCL) tailored for preserving details while retaining context window to address the unique challenges from large image size and small objects. Then, we propose to fuse features from the FM-PaCL backbone and a frozen YOLO11n backbone in the fine-tuning stage for the detection task to also utilize upstream representation learned from the COCO dataset. Our proposed SCLA framework improves the detection performance by 8.9% in terms of mAP50 on the NVD dataset against the Yolo11n baseline. Code is available at https://github.com/Boyinglby/Sideload-Contrastive-Learning-Adaptation.

Chemical Exchange Saturation Transfer (CEST) MRI enables noninvasive molecular imaging by selectively saturating exchangeable protons. Conventional tissue-specific CEST analysis often relies on projecting segmentations derived from structural MRI onto the CEST plane, introducing dependence on anatomical priors and additional registration steps. This proof-of-concept study investigates whether voxel-wise brain tissue identity can be inferred directly from CEST Z-spectral features, allowing tissue maps to be reconstructed in the native CEST acquisition space without anatomical input during inference. Eight classifiers spanning classical machine-learning methods and deep-learning architectures were trained and evaluated using 45-point voxel-wise spectra derived from B0-corrected, normalized, and spline-interpolated two-dimensional CEST data. Data were collected from six healthy subjects with 72 scans. Five subjects were used for subject-wise leave-one-subject-out cross-validation, while the sixth was reserved for independent held-out testing. Reference tissue labels for white matter, gray matter, and cerebrospinal fluid were generated from T1-weighted segmentations using FSL FAST and SPM12 and registered to the CEST plane. The evaluated classifiers comprised a one-dimensional Transformer and seven comparator models: SVM, Random Forest, FCNN, CNN, TCN-LSTM, ResNet, and Inception. On the held-out subject, the Transformer achieved a mean 88.48

This intervention study investigates the use of counterarguments in writing for critical thinking by students in the context of Generative AI (GenAI). This is especially as risks of cheating and cognitive offloading exist with the use of GenAI. We presented 36 students in a particular university course with 4 carefully selected thesis statements (from a set of popular debates) to write about anyone of them. We used six established rubrics (focus, logic, content, style, correctness and reference) to conduct three human assessments (two student peer-reviews and one experienced teacher) per writeup on a 5-point Likert scale for all the qualified samples (n) of 35 submissions (after disqualifying one for irregularity). Using the same rubrics and guidelines, we also assessed the submissions using six frontier LLMs as judges. Our mixed-method design included qualitative open-ended feedback per assessment and quantitative methods. The results reveal that (1) the students' self-written counterarguments to AI-generated content contains logic, among other things, which is a key component of critical thinking, and (2) GenAI can be successfully used at scale to assess students' written work, based on clear rubrics, and these assessments generally align with human assessments as shown with Gwets AC2 inter-rater reliability values of 0.33 for all the models except one.

The convergence of machine learning (ML) and Internet of Things (IoT) technologies has catalyzed the rise of edge intelligence, where computation is brought closer to data sources to enable real-time, autonomous decision-making. However, deploying ML at the network edge introduces significant challenges due to edge devices’ constrained computational, storage, and communication resources. This paper presents a structured review of the major challenges associated with enabling machine learning at the edge in IoT environments. The challenges are categorized into three main domains: learning limitations at the edge, hardware and resource constraints, and communication and synchronization challenges in distributed edge systems. In addition, the paper highlights key research gaps that currently limit the practical deployment of scalable edge intelligence solutions, including limited model diversity for resource-constrained environments, insufficient adoption of hybrid edge–cloud architectures, and the need for more robust distributed learning frameworks such as federated learning. By synthesizing these challenges and open research directions, this work aims to guide future research toward more scalable, privacy-aware, and intelligent edge-enabled IoT systems.

Model-Agnostic Meta-learning (MAML) is a widely adopted few-shot learning (FSL) method designed to mitigate the dependency on large, labeled datasets of deep learning-based methods in medical imaging analysis. However, MAML's reliance on a fixed number of gradient descent (GD) steps for task adaptation results in computational inefficiency and task-level overfitting. To address this issue, we introduce Tra-MAML, which optimizes the balance between model adaptation capacity and computational efficiency through a trapezoidal step scheduler (TRA). The TRA scheduler dynamically adjusts the number of GD steps in the inner optimization loop: initially increasing the steps uniformly to reduce variance, maintaining the maximum number of steps to enhance adaptation capacity, and finally decreasing the steps uniformly to mitigate overfitting. Our evaluation of TraMAML against selected FSL methods across four medical imaging datasets demonstrates its superior performance. Notably, Tra-MAML outperforms MAML by 13.36% on the BreaKHis40X dataset in the 3-way 10-shot scenario.

With the growing demand for raw materials, there is also an increased need in faster and more efficient processes of mineral exploration, especially locating possible materials to mine. Machine learning (ML) for supporting the labor-intensive and interpretive process of drill core logging becomes increasingly relevant to address these needs, since drill cores are the most direct and physically preserved record of subsurface geology available during mineral exploration. This paper reviews the current state of the art in ML-based drill core analysis, including its capabilities, limitations, and specific challenges related to generalization and practical deployment within geological workflows. The review focuses specifically on photographic images of drill core, which have been routinely used in mineral exploration for decades. This paper presents several major contributions: It offers a structured overview of current methods, organized around three key geological tasks, which are lithology prediction, geotechnical analysis, and mineralogical prediction. Additionally, it identifies potential research gaps and proposes directions for future work, concluding with an emphasis on advancing context-aware machine learning in drill core analysis through a human-in-the-loop approach.

One of the fastest-growing domains in AI is healthcare. Given its importance, it has been the interest of many researchers to deploy ML models into the ever-demanding healthcare domain to aid doctors and increase accessibility. Delivering reliable models, however, demands a sizable amount of data, and the recent COVID-19 pandemic served as a reminder of the rampant and scary nature of healthcare that makes training models difficult. To alleviate such scarcity, many published works attempted to synthesize radiological cough data to train better COVID-19 detection models on the respective radiological data. To accommodate the time sensitivity expected during a pandemic, this work focuses on detecting COVID-19 through coughs using synthetic data to improve the accuracy of the classifier. The work begins by training a CNN on a balanced subset of the Coughvid dataset, establishing a baseline classification test accuracy of 72

In this work, we proposed a novel approach for identifying quantum phase transitions in one-dimensional quantum many-body systems using AutoEncoder (AE), an unsupervised machine learning technique, with minimal prior knowledge. The training of the AEs is done with reduced density matrix (RDM) data obtained by Exact Diagonalization (ED) across the entire range of the driving parameter and thus no prior knowledge of the phase diagram is required. With this method, we successfully detect the phase transitions in a wide range of models with multiple phase transitions of different types, including the topological and the Berezinskii-Kosterlitz-Thouless transitions by tracking the changes in the reconstruction loss of the AE. The learned representation of the AE is used to characterize the physical phenomena underlying different quantum phases. Our methodology demonstrates a new approach to studying quantum phase transitions with minimal knowledge, small amount of needed data, and produces compressed representations of the quantum states.

We thoroughly investigated the generalizability of deep learning models trained on electroencephalography (EEG) data to detect Alzheimer's disease and dementia at the individual subject level. Although average model performance appears strong, it may obscure large inter-individual variability, raising concerns for clinical deployment. We trained a Hopfield-enhanced deep neural network on a publicly available EEG dataset consisting of 88 participants, including individuals diagnosed with Alzheimer's disease (AD), frontotemporal dementia (FTD), and cognitively normal controls (CN). Resting-state EEG recordings were segmented and used to train the model in a leave-one-subject-out (LOSO) cross-validation setup across multiple detection tasks: AD vs. CN, AD vs. FTD, FTD vs. CN, and AD vs. FTD vs. CN. While the model demonstrated high average performance (e.g., up to 83% accuracy), subject-level results revealed inconsistencies. Some individuals achieved perfect prediction even at the first training epoch, suggesting spurious memorization, while others predicted falsely throughout, with performance below chance. These patterns persisted despite consistent training conditions and no data leakage. Our findings highlight that strong group-level performance may be misleading in clinical settings, where decisions are made at the individual level. The models should be generalizable across individuals and be evaluated per individual before being considered for diagnostic use. Hopfield networks show promise in capturing patterns in EEG data, but patient-level validation and transparent reporting are essential to avoid premature clinical translation. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement Promobilia Foundation ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The datasets are already public on the OpenNeuro platform, and the dataset authors have already published a paper on that dataset; we have cited it. https://openneuro.org/datasets/ds004504/versions/1.0.8 I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes https://openneuro.org/datasets/ds004504/versions/1.0.8 Andreas Miltiadous, Katerina D. Tzimourta, Theodora Afrantou, Panagiotis Ioannidis, Nikolaos Grigoriadis, Dimitrios G. Tsalikakis, Pantelis Angelidis, Markos G. Tsipouras, Evripidis Glavas, Nikolaos Giannakeas, and Alexandros T. Tzallas (2024). A dataset of EEG recordings from: Alzheimer's disease, Frontotemporal dementia and Healthy subjects. OpenNeuro. [Dataset] doi: doi:10.18112/openneuro.ds004504.v1.0.8

Lithium-ion batteries are central to contemporary energy storage systems, yet the precise estimation of critical states—state of charge (SOC), state of health (SOH), and remaining useful life (RUL)—remains a complex challenge under dynamic and varied conditions. Conventional methodologies often fail to meet the required adaptability and precision, leading to a growing emphasis on the application of machine learning (ML) techniques to enhance battery management systems (BMS). This review examines a decade of progress (2013–2024) in ML-based state estimation, meticulously analysing 58 pivotal publications selected from an initial corpus of 2414 studies. Unlike existing reviews, this work uniquely emphasizes the integration of novel frameworks such as Tiny Machine Learning (TinyML) and Scientific Machine Learning (SciML), which address critical limitations by offering resource-efficient and interpretable solutions. Through detailed comparative analyses, the review explores the strengths, weaknesses, and practical considerations of various ML methodologies, focusing on trade-offs in computational complexity, real-time implementation, and generalization across diverse datasets. Persistent barriers, including the absence of standardized datasets, stagnation in innovation, and scalability constraints, are identified alongside targeted recommendations. By synthesizing past advancements and proposing forward-thinking approaches, this review provides valuable insights and actionable strategies to drive the development of robust, scalable, and efficient energy storage technologies.

This study evaluates feature selection using ANOVA and Linear Regression to optimize GRU-based models for predicting foot position in powered prostheses across varied terrains. Kinematic data from ten healthy participants during walking, stair ascend/descend, and standing were processed in MATLAB. Selected features, compared with Recursive Feature Elimination, trained GRU networks on mixed datasets and were tested on independent subjects. Results showed ANOVA and regression efficiently selected features with reduced computation and comparable performance. The GRU achieved RMSE as low as 0.066 radians, demonstrating robust generalization. While promising, clinical validation on amputee subjects remains necessary.

Colorectal cancer (CRC) poses a significant global health burden, where early and accurate diagnosis is vital to improving patient outcomes. However, the structural complexity of CRC histopathological images renders manual analysis time-consuming and error-prone. This study aims to develop an automated deep learning framework that enhances classification accuracy and efficiency in CRC diagnosis. The proposed model integrates domain-specific transfer learning and multi-model feature fusion to address challenges such as multi-scale structures, noisy labels, class imbalance, and fine-grained subtype classification. The model first applies domain-specific transfer learning to extract highly relevant features from histopathological images. A multi-head self-attention mechanism then fuses features from multiple pre-trained models, followed by a multilayer perceptron (MLP) classifier for final prediction. The framework was evaluated on three publicly available CRC datasets: EBHI, Chaoyang, and COAD. The model achieved a classification accuracy of 99.68% on the EBHI dataset (200 × subset), 86.72% on the Chaoyang dataset, and 99.44% on the COAD dataset. These results demonstrate strong generalization across diverse and complex histopathological image conditions. This study highlights the effectiveness of combining domain-specific transfer learning with multi-model feature fusion and attention mechanisms for CRC classification. The proposed model offers a reliable and efficient tool to support pathologists in diagnostic workflows, with the potential to reduce manual workload and improve diagnostic consistency.

This article presents an intervention study on the effects of the combined methods of 1) the Socratic method, 2) chain-of-thought (CoT) reasoning, 3) simplified gamification, and 4) formative feedback on university students’ math learning driven by large language models (LLMs). We call our approach Mathematics Explanations through Games by AI LLMs (MEGA). Some students struggle with math, and as a result, avoid math-related disciplines or subjects despite the importance of math across many fields, including signal processing. Oftentimes, students’ math difficulties stem from suboptimal pedagogy. We compared the MEGA method to the traditional step-by-step (CoT) method to ascertain which is better by using a within-group design after randomly assigning questions for the participants, who are university students. Samples ${(}{n}{=}{60}{)}$ were randomly drawn from each of the two test sets of the Grade School Math 8 K (GSM8K) and Mathematics Aptitude Test of Heuristics (MATH) datasets, based on an error margin of 11%, a confidence level of 90%, and a manageable number of samples for the student evaluators. These samples were used to evaluate two capable LLMs at length [Generative Pretrained Transformer 4o (GPT4o) and Claude 3.5 Sonnet] out of the initial six that were tested for capability. The results showed that students agree in more instances that the MEGA method is experienced as better for learning for both datasets. It is even much better than the CoT (47.5% compared to 26.67%) in the more difficult MATH dataset, indicating that MEGA is better at explaining difficult math problems. We also calculated the accuracies of the two LLMs and showed that model accuracies differ for the methods. MEGA appears to expose the hallucination challenge that still exists with these LLMs better than CoT. We provide public access to the MEGA app, the preset instructions that we created, and the annotations by the students for transparency.

In the world of autonomous systems and aerial surveillance, the quest to efficiently detect vehicles in diverse environmental conditions has emerged as a pivotal challenge. While these technologies have made significant advancements in the identification of objects under ordinary circumstances, the complexities introduced by snow-laden landscapes present a unique set of hurdles. The deployment of unmanned aerial vehicles (UAVs) equipped with state-of-the-art detectors in snowy regions has become an area of intense research, as it holds promise for various applications, from search and rescue operations to efficient transportation management. This paper explores the complexities that surface when it comes to identifying vehicles within snowy landscapes through the utilization of drones. It delves into the intricacies of this state-ofthe-art undertaking, offering insights into potential future directions to tackle these challenges for the unique demands of such environments. The research aims to apply the conventional procedures typically used to enhance the performance of stateof-the-art (STOA) detectors such as YOLO and faster RCNN. This is done to underscore that adhering to traditional approaches may not suffice to achieve the desired level of efficiency and accuracy when viewed from an industrial standpoint. The code and the dataset will be available at https://nvd.ltu-ai.dev/

This comprehensive systematic review critically analyzes the current progress and challenges in automating transtibial prosthesis alignment. The manual identification of alignment changes in prostheses has been found to lack reliability, necessitating the development of automated processes. Through a rigorous systematic search across major electronic databases, this review includes the highly relevant studies out of an initial pool of 2111 records. The findings highlight the urgent need for automated alignment systems in individuals with transtibial amputation. The selected studies represent cutting-edge research, employing diverse approaches such as advanced machine learning algorithms and innovative alignment tools, to automate the detection and adjustment of prosthesis alignment. Collectively, this review emphasizes the immense potential of automated transtibial prosthesis alignment systems to enhance alignment accuracy and significantly reduce human error. Furthermore, it identifies important limitations in the reviewed studies, serving as a catalyst for future research to address these gaps and explore alternative machine learning algorithms. The insights derived from this systematic review provide valuable guidance for researchers, clinicians, and developers aiming to propel the field of automated transtibial prosthesis alignment forward.