Deep neural networks such as convolutional neural networks (CNNs) and transformers have achieved many successes in image classification in recent years. It has been consistently demonstrated that best practice for image classification is when large deep models can be trained on abundant labelled data. However there are many real world scenarios where the requirement for large amounts of training data to get the best performance cannot be met. In these scenarios transfer learning can help improve performance. To date there have been no surveys that comprehensively review deep transfer learning as it relates to image classification overall. However, several recent general surveys of deep transfer learning and ones that relate to particular specialised target image classification tasks have been published. We believe it is important for the future progress in the field that all current knowledge is collated and the overarching patterns analysed and discussed. In this survey we formally define deep transfer learning and the problem it attempts to solve in relation to image classification. We survey the current state of the field and identify where recent progress has been made. We show where the gaps in current knowledge are and make suggestions for how to progress the field to fill in these knowledge gaps. We present a new taxonomy of the applications of transfer learning for image classification. This taxonomy makes it easier to see overarching patterns of where transfer learning has been effective and, where it has failed to fulfill its potential. This also allows us to suggest where the problems lie and how it could be used more effectively. We show that under this new taxonomy, many of the applications where transfer learning has been shown to be ineffective or even hinder performance are to be expected when taking into account the source and target datasets and the techniques used.

Large language model agents increasingly solve complex tasks by composing reusable skills from a library. To address this, the key challenge is not merely to retrieve individually relevant skills, but to identify a complete and executable skill composition. In this paper, we argue that this problem can be solved in a graph with three levels: compositional relations among skill queries, similarity between queries and candidates in the skill library, and the dependencies among the selected candidates. We introduce SkillTrace, which organizes the user query into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies. Experiments on SkillsBench and ALFWorld demonstrate that SkillTrace achieves state-of-the-art performance, reaching a success rate of 53.17

Large language models (LLMs) are becoming an increasingly important component of human–computer interaction, enabling users to coordinate a wide range of intelligent agents through natural language. While language-based interfaces are powerful and flexible, they implicitly assume that users can reliably produce explicit linguistic input, an assumption that may not hold for users with speech or motor impairments, e.g., Amyotrophic Lateral Sclerosis (ALS). In this work, we investigate whether neural signals can be used as an alternative input to LLMs, particularly to support those socially marginalized or underserved users. We build a simple brain-LLM interface, which uses EEG signals to guide image generation models at test time. Specifically, we first train a classifier to estimate user satisfaction from EEG signals. Its predictions are then incorporated into a test-time scaling (TTS) framework that dynamically adapts model inference using neural feedback collected during user evaluation. The experiments show that EEG can predict user satisfaction, suggesting that neural activity carries information on real-time preference inference. These findings provide a first step toward integrating neural feedback into adaptive language-model inference, and hopefully open up new possibilities for future research on adaptive LLM interaction.

Test-time scaling offers a promising way to improve the reasoning performance of vision-language large models (VLLMs) without additional training. In this paper, we explore a simple but effective approach for applying test-time scaling to chest X-ray report generation. Specifically, we introduce a lightweight Thought Graph Traversal (TGT) framework that guides the model to reason through organ-specific findings in a medically coherent order. This framework integrates structured medical priors into the prompt, enabling deeper and more logical analysis with no changes to the underlying model. To further enhance reasoning depth, we apply a reasoning budget forcing strategy that adjusts the model's inference depth at test time by dynamically extending its generation process. This simple yet powerful combination allows a frozen radiology VLLM to self-correct and generate more accurate, consistent chest X-ray reports. Our method outperforms baseline prompting approaches on standard benchmarks, and also reveals dataset biases through traceable reasoning paths. Code and prompts are open-sourced for reproducibility at https://github.com/glerium/Thought-Graph-Traversal

The growth of networked and IoT systems has intensified cyber-security threats and exposed the limits of traditional signature-based intrusion detection. Although machine-learning-based intrusion detection systems often report strong benchmark performance, high ac- curacy within a single dataset does not necessarily guarantee reliable performance in unseen network environments. This study investigates the generalisation capability of supervised machine learning models for intrusion detection using UNSW-NB15 and TON_IoT. Random Forest, Logistic Regression, and Naive Bayes were evaluated under same-dataset and cross-dataset settings. Random Forest achieved the strongest same dataset performance, with 95.08

Robot teammates could re-embody to enhance communication through gaze and gestures in distributed human-robot teaming (HRT) within virtual reality (VR) environments, offering potential benefits for time-critical domains such as emergency response. Yet despite this promise, robot re-embodiment remains underexplored in immersive settings and there is little clarity on how re-embodied robots should be visually designed. In this mixed-methods study, we used Immersive Speculative Enactments to investigate how people make sense of re-embodying robot teammates based on their visual forms in a shared VR environment. In pairs, 42 participants enacted three emergency scenarios with speculative robot teammates (a drone, humanoid firefighter, and fire truck) presented in machine-like, augmented, and human-like embodiments. Thematic analysis of post-session interviews, supported by quantitative data on participants’ VR experiences, showed that sense-making was shaped by robot type, task context, and social norms. Machinelike forms supported recognition of identity and function and were generally preferred; augmented forms blended social and functional cues; and humanlike forms prompted richer social interpretations but sometimes blurred identity and role boundaries. Participants’ future imaginaries of re-embodiment reflected concerns about contextual risk, automation reliability, and emotional labour. Together, these findings provide early insights into how identity, form, and function shape the interpretation of robot teammates in VR, offering guidance for the design of future re-embodiment systems and directions for further research.

Large-scale radiology data are critical for developing robust medical AI systems. However, sharing such data across hospitals remains heavily constrained by privacy concerns. Existing de-identification research in radiology mainly focus on removing identifiable information to enable compliant data release. Yet whether de-identified radiology data can still preserve sufficient utility for large-scale vision-language model training and cross-hospital transfer remains underexplored. In this paper, we introduce a utility-preserving de-identification pipeline (UPDP) for cross-hospital radiology data sharing. Specifically, we compile a blacklist of privacy-sensitive terms and a whitelist of pathology-related terms. For radiology images, we use a generative filtering mechanism that synthesis a privacy-filtered and pathology-reserved counterparts of the original images. These synthetic image counterparts, together with ID-filtered reports, can then be securely shared across hospitals for downstream model development and evaluation. Experiments on public chest X-ray benchmarks demonstrate that our method effectively removes privacy-sensitive information while preserving diagnostically relevant pathology cues. Models trained on the de-identified data maintain competitive diagnostic accuracy compared with those trained on the original data, while exhibiting a marked decline in identity-related accuracy, confirming effective privacy protection. In the cross-hospital setting, we further show that de-identified data can be combined with local data to yield better performance.

Misinformation poses significant risks to public opinion, health, and security. While most fake news detection methods rely on text analysis, little is known about how people physically respond to false information or repeated exposure to the same statements. This study investigates whether wearable sensors can detect belief in a statement or prior exposure to it. We conducted a controlled experiment where participants evaluated statements while wearing an EmotiBit sensor that measured their skin conductance (electrodermal activity, EDA) and peripheral blood flow (photoplethysmography, PPG). From 28 participants, we collected a dataset of 671 trials, each labeled with whether the participant believed the statement and whether they had seen it before. This dataset introduces a new resource for studying physiological responses to misinformation. Using machine learning models, including KNN, CNN, and LightGBM, we analyzed these physiological patterns. The best-performing model achieved an accuracy of 62.69% on the belief classification task, with EDA outperforming PPG. These findings demonstrate the potential of wearable sensors as a minimally intrusive tool for detecting belief and prior exposure, offering new directions for real-time misinformation detection and adaptive, user-aware systems.

Large language models (LLMs) have revolutionised many fields, with LLM-as-a-service (LLMSaaS) offering accessible, general-purpose solutions without costly task-specific training. In contrast to the widely studied prompt engineering for directly solving tasks (in vivo), this paper explores LLMs' potential for in-vitro applications: using LLM-generated labels to improve supervised training of mainstream models. We examine two strategies - (1) noisy label correction and (2) training data augmentation - in empathy computing, an emerging task to predict psychology-based questionnaire outcomes from inputs like textual narratives. Crowdsourced datasets in this domain often suffer from noisy labels that misrepresent underlying empathy. We show that replacing or supplementing these crowdsourced labels with LLM-generated labels, developed using psychology-based scale-aware prompts, achieves statistically significant accuracy improvements. Notably, the RoBERTa pre-trained language model (PLM) trained with noise-reduced labels yields a state-of-the-art Pearson correlation coefficient of 0.648 on the public NewsEmp benchmarks. This paper further analyses evaluation metric selection and demographic biases to help guide the future development of more equitable empathy computing models.

We explore a situation in which the target domain is accessible, but real-time data annotation is not feasible. Instead, we would like to construct an alternative training set from a large-scale data server so that a competitive model can be obtained. For this problem, because the target domain usually exhibits distinct modes (i.e., semantic clusters representing data distribution), if the training set does not contain these target modes, the model performance would be compromised. While prior existing works improve algorithms iteratively, our research explores the often-overlooked potential of optimizing the structure of the data server. Inspired by the hierarchical nature of web search engines, we introduce a hierarchical data server, together with a bipartite mode matching algorithm (BMM) to align source and target modes. For each target mode, we look in the server data tree for the best mode match, which might be large or small in size. Through bipartite matching, we aim for all target modes to be optimally matched with source modes in a one-on-one fashion. Compared with existing training set search algorithms, we show that the matched server modes constitute training sets that have consistently smaller domain gaps with the target domain across object re-identification (re-ID) and detection tasks. Consequently, models trained on our searched training sets have higher accuracy than those trained otherwise. BMM allows data-centric unsupervised domain adaptation (UDA) orthogonal to existing model-centric UDA methods. By combining the BMM with existing UDA methods like pseudo-labeling, further improvement is observed.

Pain recognition using physiological signals is a critical step toward advancing automatic pain assessment, particularly in clinical and experimental settings. In this study, we present a machine learning-based approach to address the AI4PAIN 2025 challenge, which involves classifying pain intensity into three levels—No Pain, Low Pain, and High Pain—based on multimodal time-series signals including electrodermal activity (EDA), blood volume pulse (BVP), respiration (RESP), and peripheral oxygen saturation (SpO2). After extracting relevant statistical and frequency-domain features from each modality, we evaluated a series of classical machine learning models: RandomForestClassifier, LGBMClassifier, HistGradientBoostingClassifier, XGBClassifier, and CatBoostClassifier. Among these, the Random Forest Classifier emerged as the most robust model, achieving the highest test accuracy of 62.07%, indicating better generalization compared to others. Although the XGBClassifier attained the highest validation accuracy of 68%, its performance degraded more noticeably on the test set, suggesting potential overfitting. Both LGBMClassifier and HistGradientBoostingClassifier reached similar validation accuracies (around 64%), with the latter showing better class-wise balance under label imbalance conditions. However, Random Forest demonstrated the most consistent performance across both validation and test evaluations, making it the most reliable model in this study. The entire preprocessing, feature engineering, and classification pipeline developed for this task is encapsulated in our proposed system, Painthenticate, designed to support reproducible and interpretable pain classification. These findings highlight the value of model selection not only based on aggregate accuracy but also on consistency across class distributions in physiological pain classification tasks.

Our workshop brings together domain experts and research students to share insights, practical guidance, and evaluations on key topics, including social network analysis, graph algorithms, web mining, semantics and knowledge, security, privacy, fairness, and ethics on the web. We invite survey, evaluation, or review papers that critically analyze models and datasets from diverse perspectives. These papers serve as essential resources by (i) providing quick reference guides for researchers and practitioners, (ii) enhancing accessibility for newcomers, and (iii) distilling key insights into actionable knowledge. Complementing these contributions, invited talks from experts and industry leaders will offer practical perspectives, fostering cross-domain collaboration in web technologies. Through thought-provoking discussions and networking opportunities, the workshop bridges research and real-world applications, setting a new standard for interdisciplinary exchange in the field.

Can the relative performance of a pre-trained large multimodal model (LMM) be predicted without access to labels? As LMMs proliferate, it becomes increasingly important to develop efficient ways to choose between them when faced with new data or tasks. The usual approach does the equivalent of giving the models an exam and marking them. We opt to avoid marking and the associated labor of determining the ground-truth answers. Instead, we explore other signals elicited and ascertain how well the models know their own limits, evaluating the effectiveness of these signals at unsupervised model ranking. We evaluate 47 state-of-the-art LMMs (e.g., LLaVA) across 9 visual question answering benchmarks, analyzing how well uncertainty-based metrics can predict relative model performance. Our findings show that uncertainty scores derived from softmax distributions provide a robust and consistent basis for ranking models across various tasks. This facilitates the ranking of LMMs on unlabeled data, providing a practical approach for selecting models for diverse target domains without requiring manual annotation.

Robotic systems have been increasingly applied across a wide range of sectors to alleviate the burden on human labor, enhancing efficiency in various work settings. In certain instances, effective human-robot collaboration is essential for task success, and it can be facilitated by equipping robots with emotional and empathetic capabilities that resemble those of humans. Although interpreting emotions via facial cues or physical gestures is comparatively accessible, replicating and modeling empathy computationally poses a significantly greater challenge. This paper outlines the architecture and performances of computational models developed for simulating and predicting human empathy. The OMG-Empathy dataset, comprising storytelling video recordings and self-reported empathy valence ratings from listeners after viewing, was used to train models aimed at predicting empathy responses of listeners. A range of multimodal features, such as conversational content, facial expressions, and vocal arousal, were extracted for analysis. These features were examined to assess their impact on the prediction of empathy valence. For computational models, various Machine Learning (ML) techniques were implemented, namely Support Vector Machines (SVM), Decision Trees, Random Forests, and Neural Networks (NN), with the Concordance Correlation Coefficient (CCC) employed as the primary evaluation criterion. Among all models evaluated, the Support Vector Machine (SVM) achieved the greatest performance, yielding the highest personalized and generalized CCC scores of around 0.07. Result also indicated that linguistic empathy feature had a stronger correlation with overall empathy than the emotional empathy feature. This research offers insights on computational modeling on empathy and a potential inspiration to develop more human-like robots. Future work may involve more refining empathy extraction approaches and more robust models which are adaptable to a wider range of cases.

In computer vision tasks, features often come from diverse representations, domains (e.g., indoor and outdoor), and modalities (e.g., text, images, and videos). Effectively fusing these features is essential for robust performance, especially with the availability of powerful pre-trained models like vision-language models. However, common fusion methods, such as concatenation, element-wise operations, and non-linear techniques, often fail to capture structural relationships, deep feature interactions, and suffer from inefficiency or misalignment of features across domains or modalities. In this paper, we shift from high-dimensional feature space to a lower-dimensional, interpretable graph space by constructing relationship graphs that encode feature relationships at different levels, e.g., clip, frame, patch, token, etc. To capture deeper interactions, we expand graphs through iterative graph relationship updates and introduce a learnable graph fusion operator to integrate these expanded relationships for more effective fusion. Our approach is relationship-centric, operates in a homogeneous space, and is mathematically principled, resembling element-wise relationship score aggregation via multilinear polynomials. We demonstrate the effectiveness of our graph-based fusion method on video anomaly detection, showing strong performance across multi-representational, multi-modal, and multi-domain feature fusion tasks.

Empirical study designs in HCI evolve in response to temporal realities and technological advancements. In this context, virtual reality (VR) shows potential for new empirical research designs, going beyond the (still) quite dominantly lab-based research roots in HCI. Previous work has been conducted to identify the use of VR for gathering non-homogeneous and representative sample populations and for conducting empirical studies in resource-constrained environments. Yet, it is unclear how VR empirical user study designs affect the participants' behavior and experience, potentially influencing the study results compared to in-situ/in-lab studies. In this paper, we conducted a gesture elicitation study (GES) in a realistic physical smart room and its digital duplicate in VR. Sixty-six participants' responses were collected using standardized questionnaires along with between-group gesture agreement analysis. Our comparison shows that the VR study produces a higher number of unique gesture proposals and similar best gestures to the in-person study for 95.4% of the referents, with minimum influence on the gesture proposals. We further discuss the usability, pragmatic and hedonic qualities, presence, task load, and implications of using VR for GESs, and highlight future directions for using VR-based empirical study designs. We found that VR can produce reliable data and improve participant experience with the same task load, making it viable to conduct remote GES and a substitute for conventional lab-based experiments.

Detecting deceptive behaviour for surveillance and border protection is critical for a country’s security. With the advancement of technology in relation to sensors and artificial intelligence, recognising deceptive behaviour could be performed automatically. Following the success of affective computing in emotion recognition from verbal and nonverbal cues, we aim to apply a similar concept for deception detection. Recognising deceptive behaviour has been attempted; however, only a few studies have analysed this behaviour from gait and body movement. This research involves a multimodal approach for deception detection from gait, where we fuse features extracted from body movement behaviours from a video signal, acoustic features from walking steps from an audio signal, and the dynamics of walking movement using an accelerometer sensor. Using the video recording of walking from the Whodunnit deception dataset, which contains 49 subjects performing scenarios that elicit deceptive behaviour, we conduct multimodal two-category (guilty/not guilty) subject-independent classification. The classification results obtained reached an accuracy of up to 88% through feature fusion, with an average of 60% from both single and multimodal signals. Analysing body movement using single modality showed that the visual signal had the highest performance followed by the accelerometer and acoustic signals. Several fusion techniques were explored, including early, late, and hybrid fusion, where hybrid fusion not only achieved the highest classification results, but also increased the confidence of the results. Moreover, using a systematic framework for selecting the most distinguishing features of guilty gait behaviour, we were able to interpret the performance of our models. From these baseline results, we can conclude that pattern recognition techniques could help in characterising deceptive behaviour, where future work will focus on exploring the tuning and enhancement of the results and techniques.