This study examines how Large Language Models (LLMs) can reduce biases in text-to-image generation systems by modifying user prompts. We define bias as a model's unfair deviation from population statistics given neutral prompts. Our experiments with Stable Diffusion XL, 3.5 and Flux demonstrate that LLM-modified prompts significantly increase image diversity and reduce bias without the need to change the image generators themselves. While occasionally producing results that diverge from original user intent for elaborate prompts, this approach generally provides more varied interpretations of underspecified requests rather than superficial variations. The method works particularly well for less advanced image generators, though limitations persist for certain contexts like disability representation. All prompts and generated images are available at https://iisys-hof.github.io/llm-prompt-img-gen/
Kolmogorov-Arnold Networks (KANs) represent a breakthrough in deep learning, diverging from Multi-Layer Perceptrons (MLPs) by generalizing the Kolmogorov-Arnold representation theorem (KAT) to networks of arbitrary depth and width. This theorem facilitates the decomposition of multivariate functions into constituent one-dimensional elements, with learnable activation functions on weights and the sum operator on nodes. KANs have been shown to exhibit robust performance in function approximation, validated across mathematical, physical, and practical domains such as traffic prediction and medical diagnostics. Our study investigates KANs' efficacy through comprehensive evaluations on OpenML, Kaggle and UCI datasets, with a focus on enhancing Human Activity Recognition systems. They demonstrate high classification performance compared to conventional machine learning approaches and MLPs. These findings underscore KANs' potential as scalable, interpretable tools in modern machine learning applications given their favorable neural scaling laws.
This study investigates the application of Hierarchical clustering for image segmentation, with a focus on its efficacy in analyzing medical images, particularly MRI scans of brain tumors. Image segmentation plays a pivotal role in computer vision, facilitating various applications across industries. Leveraging a systematic approach, we conduct a comprehensive review of recent literature on machine learning algorithms for image segmentation. Subsequently, utilizing a dataset comprising MRI images with and without tumors, we preprocess and analyze the data using the Histogram of Oriented Gradients (HOG) technique to extract pertinent features. These features serve as input for the Hierarchical clustering algorithm to partition the images into distinct regions of interest. For each row of vectors, the Jensen-Shenton distance was calculated. The resulting symmetric matrices are distances among the corresponding vectors, quantifying dissimilarity in cluster analysis. Our findings underscore the effectiveness of Hierarchical clustering in clustering medical images, with potential implications for advancing computational analysis in healthcare and related domains.
This work addresses the challenge of transferring large-scale deep learning vision models to domain-specific applications, where models pre-trained on generic datasets often lack the nuanced representations necessary for accurately capturing subtle context variations. We investigate the efficacy of pre-training Vision Transformers (ViTs) on domain-specific synthetic datasets generated via Stable Diffusion, compared to traditional pre-training on off-the-shelf generic datasets. We created three synthetic datasets—landscapes, fruits and vegetables, and birds—each containing 500,000 images. For each dataset, we pre-trained ViT models on subsets of 200K, and 500K images, and fine-tuned these models on two corresponding downstream classification tasks. Additionally, we pre-trained ViT models on the full ImageNet, with 1.2M images, and subsets of 200K, 500K images, and fine-tuned them on the same downstream tasks. Results indicate that ViT models pre-trained on the synthetic datasets (landscapes, fruits and vegetables, and birds) outperformed those pre-trained on generic datasets in terms of classification accuracy on domain specific tasks. These findings underscore the potential of using synthetic datasets tailored to specific domains for enhancing the performance of vision transformers in specialized tasks.
Network intrusion, which involves unauthorized access to computer systems, poses a substantial threat in today’s digital environment. This threat manifests through various tactics, such as malware, phishing, and DoS attacks. These attacks can lead to significant financial and reputational damage, data breaches, and operational disruptions. Consequently, the ability to identify and respond to network attacks is critical for safeguarding against malicious online activities. The objective of our research is to investigate the efficacy of machine learning algorithms in detecting such attacks, utilizing the UNSW-NB15 dataset. The dataset includes two distinct classification targets: categorizing different types of attacks and distinguishing between malicious attacks and normal (benign) activities. The primary goal of our study is to determine the most effective classifier for these tasks by addressing specific challenges in the dataset, overlapping data, and class imbalance. These issues can significantly impact the performance of machine learning models if they are not properly managed. To tackle these challenges, the first phase of our approach is to apply class decomposition and oversampling. Class decomposition: apply clustering algorithms to the majority class to reduce its dominance over minority classes. Oversampling helps to balance the dataset. The second phase is to find the best model and the best combination of hyperparameters that boost the classification performance to the max. Through our experimentation, we have found that the best-performing models are: Multi-Layer Perceptron (MLP) achieved a remarkable 99.04% F1-score for attack detection, highlighting its precision and recall in identifying malicious activities. Similarly, the XGBOOST algorithm achieved a 98.89% F1-score for multiclass attack categorization, showcasing its effectiveness in differentiating between various types of network attacks. Both models demonstrate robust performance (F1-score) across all classes in accurately identifying and categorizing different types of network intrusions. These results suggest that machine learning, when applied with appropriate techniques to address dataset challenges, holds significant promise for enhancing cybersecurity measures and protecting digital environments from evolving threats.
This study investigates the challenges of permeability prediction in reservoir engineering, focusing on addressing uncertainties inherent in the data and modelling process, and leveraging Nuclear Magnetic Resonance (NMR) log data from the Northern Sea Volve field. The study uses a probabilistic machine learning method called Gaussian Process Regression (GPR) with different kernels, such as Matern52, Matern32, and Radial Basis Function (RBF). LSboost, K-nearest neighbour (KNN), and XGBoost are some of the existing models that are used for comparison. Performance metrics including Mean Absolute Error (MAE), Mean Squared Error (MSE), and coefficient of determination (R2) are utilized for assessment. Additionally, the uncertainty associated with different GPR kernels is analyzed, and confidence intervals are generated to provide insights into model behaviour. The inclusion of confidence intervals enhances interpretability by quantifying the range within which the true permeability value is likely to fall with a specified level of confidence, offering valuable information for decision-making processes in reservoir engineering applications. Findings demonstrate the effectiveness of GPR with Matern52 and Matern32 kernels in permeability prediction, offering competitive performance and robust uncertainty quantification. This research contributes to advancing reservoir engineering by providing a comprehensive and uncertainty-aware approach to permeability prediction.
This study focuses on developing a custom Speech-to-Text (STT) application to improve data collection from technicians working in a cleanroom environment. Comment data is traditionally captured via keyboard input. However, within a cleanroom, employees wear gloves making typing difficult, they may not have free hands and workstations can be located far from the tools requiring travel. As a result comments are often short and lack information, the length for a one hour job and a six hour job are often comparable so, there is significant information loss. Industries often have their own specific language and existing STT systems are not able to accurately capture this, so a custom model is required. The small.en Whisper AI Model was taken as a starting point and refined using a training dataset of recordings taken from the application environment. A PoC implementation of the STT capability was developed using an Android application deployed on a voice controlled Smart Helmet with Private 5G Network and Wifi connectivity that enables hands free capture of comments at the source. The study also explores the potential applications of Automated Speech Recognition systems for transcriptions and translations, and contributes to the ongoing efforts in leveraging STT technology to drive continuous improvement and innovation in industrial operations.
Social robots have emerged recently as a technology that could provide social support and increase companionship, especially for older adults. Effective human-robot interaction requires social robots to interact empathically and naturally by using Artificial Emotional Intelligence (AEI), which allows them to recognise different multimodal inputs and engage with users emotionally. To this end, we developed AEI companion pet robot classification system that incorporated various multimodal inputs, including hand gestures, emotions from images, emotions from voice annotations, and speech detection. We utilised different deep learning approaches for each mode and adopted lightweight computationally efficient models. For multimodal input detection, we extended the MediaPipe framework with a mathematical approach for hand gestures, integrated an updated MobileNet transfer learning model for facial emotions, employed a Convolutional Neural Network (CNN) model for detecting emotions from voice, and used a lightweight Whisper model as an automatic speech recognition (ASR). Our findings indicated that the modified MobileNet lightweight model achieved an accuracy of 84.60% along with an efficient gesture detection system with an accuracy of 90% integrated with an effective CNN model of 86.89% accuracy. Our results imply that developing a multimodal AEI companion pet robot could promote future research in the development of emotionally intelligent companion robots.
The rapid advancement of Artificial Intelligence (AI), particularly through Large Language Models (LLMs), has catalysed a technological revolution, leading to the widespread adoption of AI-driven chatbots across industries. OpenAI’s customisable generative pre-trained transformer (GPT) offerings have popularised generative AI, enabling organisations of all sizes to implement chatbots for customer support. This development presents an opportunity for businesses to offer 24/7, cost-efficient customer service that can overcome the historical limitations of chatbots that lack a "human element." However, despite the proliferation of AI chatbots, there remains a crucial need to evaluate their effectiveness in meeting user needs and preferences for human-like interaction. Current service quality assessment tools, such as SERVQUAL and E-SERVQUAL, are unable to evaluate AI-specific capabilities like language intelligence and recognition. Existing research also lacks information on factors that affect user satisfaction and the continued use of AI chatbots. Based on a mixed-methods study, this paper proposes a new instrument for measuring user satisfaction with AI chatbots, specifically for customer support roles. Using the Stanford five-step Design Thinking Process, this study devised a customer support AI chatbot evaluation instrument through a literature review, Cheatstorming, and SCAMPER techniques, followed by testing in a Danish company. The research employs Prentice and Nguyen’s three-stage scale development process to ensure content, reliability, and construct validity, addressing gaps in current scholarship and advancing understanding of AI chatbot user satisfaction.
Human Activity Recognition using tiny wearable devices presents unique challenges due to limited computational resources and battery life. Transformers have recently emerged as the most effective tools for time-series classification. This is due to their capacity to capture long-range dependencies and complex temporal dynamics, combined with relatively lower computational complexity, which allows them to surpass the performance of more traditional systems such as convolutional or recurrent neural networks. In fact, transformers do not process data sequentially. Instead, they leverage an attention mechanism that allows them to weigh the importance of different parts of the input data differently. This study experimentally investigates the applicability of transformer models on tiny devices by porting them on a low-power ESP32 device. Comprehensive evaluations are conducted on several benchmark datasets, demonstrating that transformer-based models are not viable solutions for tiny devices compared to convolutional and recurrent neural networks, with respect to which it achieves up to 14% lower accuracy depending on the dataset used.
Artificial intelligence (AI) technologies can potentially revolutionise society and the economy, but their use also introduces new hazards. To ensure that AI-enabled systems are trustworthy and managed responsibly, actors involved in their development, deployment, and use must act with prudence and manage risks. This paper presents aspects of the framework for assuring AI-enabled systems in the new DNV Recommended Practice (RP): DNV-RP-0671, ”Assurance of AI-enabled systems”. The objective of the RP is to guide how to assure that AI-enabled systems are trustworthy and managed responsibly throughout their entire lifecycle. The RP is based on the following key principles: stakeholder focus, evidence-based argumentation, systems approach, risk-based approach, modularity, and lifecycle perspective. The RP provides guidance on showing compliance with applicable standards and regulations, including identification and refinement of case-specific system-wide requirements down to requirements specific to the AI component of the system. Trustworthy and responsible AI involves social and technical aspects and requires a risk management system that accommodates the iterative process of assuring AI-enabled systems. Governance is also an essential aspect of assuring trustworthy and responsible AI. The framework is flexible enough to accommodate the rapid advancements in AI technology and the differing frameworks for AI governance among regions and industries.
The association between Original Equipment Manufacturers (OEMs) and sensor manufacturers must be consistently improved to accommodate the increasing usage of automotive sensors in vehicles. This requires detailed consideration of any adaptations and changes in sensor characteristics, as they directly impact the working principles of the related software applications. This study introduces a new method, the "Auto-Adaptive Diffusion (ADF) Model," which aims to uniformize the automotive camera sensor data by sensor data abstraction and, successively, reinforce the Plug-and-Play (PnP) component. The ADF model is designed to homogenize specific sensor data parameters and formulate them into a single format, reducing the interdependencies between the automotive camera sensors and their software components. It also aims to simplify the sensor configuration process, regardless of the sensor vendors, while enhancing the accuracy and reliability of applications. This work utilizes the "Drive&Act" as the Ground Truth (GT) dataset, centralizing on the In-cabin Monitoring System (IMS) with pre-trained YOLOv8 for object detection. Experimental results and comparisons show that the ADF model outperforms traditional techniques, such as Open Source Computer Vision (OpenCV) libraries and Generative Adversarial Network (GAN) based work named "Sensor-Agnostic Image Translation Framework (SAITF)" in various Key Performance Indicators (KPIs).
Bridges are pivotal components within the transportation network in all countries. However, the current state of bridges is far from optimistic—46.4% are regarded as fair, and 7.6% are deemed to be in poor condition among over 600,000 bridges in the U.S.; therefore, the accurate prediction of bridge conditions is crucial for on-time repair and maintenance. Numerous studies have leveraged efficient machine learning models to predict bridge conditions; however, the datasets related to the task are usually highly imbalanced, which may negatively impact the model’s performance. In this paper, we propose a novel framework for bridge condition prediction with machine learning models, consisting of four stages: 1) data pre-processing, 2) feature selection, 3) resampling, and 4) multi-class classification. We conduct extensive experiments across five machine learning models, including k-NN, Naïve Bayes, AdaBoost, XGBoost, and LightGBM. Among the experimented models, LightGBM achieves the best performance with SMOTE resampling, with an overall performance of nearly 100%, demonstrating its high capacity in predicting bridge conditions in real-world scenarios.
Monitoring oxygen consumption during cycling is crucial for performance and health management. Due to the time series characteristics of oxygen consumption data, it is particularly important to select a suitable model for prediction. However, despite the advantages of Gated Recurrent Unit (GRU) model in processing time series data, its application in the field of cycling oxygen consumption prediction is still limited. Therefore, this study used GRU model to predict oxygen consumption during riding, and compared it with Multi-Layer Perceptron (MLP), AutoRegressive Integrated Moving Average (ARIMA), Support Vector Machines (SVM) and Artificial Neural Networks (ANN) models. The results show that the GRU model has the best performance in forecasting accuracy and error control, and its evaluation indexes are: the Mean Absolute Error (MAE) is 0.0440, the Mean Square Error (MSE) is 0.0035, the coefficient of determination (R²) is 0.9273, and the prediction accuracy is 90.59%.
The current electricity information collection has many problems such as multiple collection types, difficulty in fault location, and low maintenance efficiency, which makes it difficult to guarantee the reliability of the electricity information collection system. Therefore, an anomaly tracing method based on generative adversarial network (GAN) and bidirectional long short-term memory neural network (BiLSTM) is proposed. First, two autoencoder networks are constructed using an encoder based on BiLSTM with multi-head attention and two decoders based on BiLSTM, which are used as generator and discriminator of GAN. Then, the generator and discriminator are adversarially trained to detect anomaly in multi-source power time series data. Finally, the matrix autoregressive model (MAR) is used to predict the detected anomalous time windows, then finish the location of abnormal system operating links. The experimental results show that the accuracy, precision, recall and F1 score of the proposed model are 0.9117, 0.8486, 0.7513 and 0.7970, respectively, which significantly outperform the state-of-the-art baselines. It is proved that the model can effectively improve the anomaly detection performance for power data, and can realize the function of anomaly monitoring and automatic localization in electricity information collection.
In the domain of machine learning and computer vision, employing deep learning models for real-time object detection tasks, especially in low-resource environments, poses significant challenges. This study aimed to assess the effectiveness and efficiency of five pre-trained neural network architectures—SqueezeNet, MobileNetV3, ResNet50, VGG16, and InceptionNet—for detecting and counting sit-ups from video footage. The evaluation criteria included accuracy in sit-up counting compared to manual methods and the computational resources utilized. Among the models examined, SqueezeNet demonstrated notable performance, closely aligning with manual counts and outperforming larger, presumably more complex networks. Despite its compact size of 3.33 MB, SqueezeNet showcased the potential for achieving high detection accuracy without the need for extensive network complexity. Conversely, larger networks such as VGG16, MobileNetV3, ResNet50, and InceptionNet exhibited significant discrepancies in sit-up counts, indicating potential overfitting during training and a lack of generalizability to real-world data. Regarding computational resources, all networks displayed minimal memory usage and CPU utilization, highlighting the feasibility of deploying them on low-resource platforms like Raspberry Pi. This study underscores the promise of lightweight deep learning models like SqueezeNet for real-time object detection tasks and provides valuable insights for future research and applications in health and fitness monitoring.
Estimating influential parameters in real-world applications can be challenging, especially in problems where the model likelihood is intractable. To perform inference, the machine learning model should be able to understand the underlying phenomena and conditions which have led to the observations in hand. Estimating the posterior distribution of the influential parameters given observations (the inverse problem) can therefore be onerous, especially when the likelihood is intractable. The Approximate Bayesian Computation (ABC) paradigm is the standard statistical methodology to address such likelihood-free problems. In a nutshell, ABC methods establish a proxy for the likelihood based on data simulations. Two main bottlenecks of any ABC technique are the comparisons between observed and simulated datasets, and the manner via which data must be summarized in order to make the eventual data comparisons optimally fit for the inference purpose. We propose an ABC algorithm that learns how to perform data summarization in a data-driven manner, which is not based on pre-selecting summary statistics a priori. The proposed ABC algorithm is based on a neural network which is trained via adversarial learning. We show how the proposed ABC algorithm can infer the posterior distributions of parameters with high fidelity, and with low bias and variance. We apply the algorithm to two different applications and demonstrate its efficacy and robustness with respect to parameter inference.
End-to-end (E2E) modeling is a promising approach that generates rich latent representations by leveraging non-verbal and acoustic information, as opposed to relying solely on linguistic data from automatically generated transcripts in cascade systems. Using advanced audio processing and clustering techniques, we propose a novel framework for unsupervised extractive meeting speech summarization. In contrast to traditional transcript-based methods, our approach generates summaries directly from raw audio recordings. Leveraging the Wav2Vec 2.0 model for feature extraction, we introduce custom segment duration-based cluster selection measures. These measures optimise key audio segments, thereby enhancing the summarization process. Our experiments on the widely-used AMI Meeting Corpus shows that our system significantly outperforms existing state-of-the-art unsupervised extractive summarization models, providing a robust and reliable solution.
Neural networks have become a cornerstone of modern machine learning, powering advances in numerous domains. In distributed computing environments, leveraging large batch sizes is crucial for the efficient training of these networks. However, this approach often leads to reduced data stochasticity when sampling information such as the loss function gradient, adversely affecting the generalization performance of the models. We tackle this issue by introducing the Sub-Batch update mechanism, which divides the large batch into sub-batches and updates the network parameters using each sub-batch. The Sub-Batch update mechanism increases the number of nonlinear operations of calculating the neural network parameter update vector from the sampled network information in Component-Wise Natural Gradient Descent (CW-NGD) neural network training algorithm. This increases the data stochasticity, resulting in accuracy improvement in the trained neural network. Experimental results show that with Sub-Batch update mechanism applied CW-NGD produces trained networks with higher accuracy compared to Adam, SGD, across all six datasets tested, including the ImageNet dastaset.
Self-powered photovoltaic sensor technology has been presented as a gestural interface that could recognise time-series data of dynamic hover gestures using machine learning (ML). To further expand and improve the classification system of dynamic hover hand gestures, we discovered time-series data from multiple subjects and explored the systems’ performance through ML algorithms. In this paper, we collected our own time-series dataset of 3,696 hand gesture samples from 48 multiple subjects using an off-the-shelf photovoltaic sensor and analysed them by applying different preprocessing techniques such as smoothing, augmentation, normalisation, and resampling. Then, we used different ML algorithms such as K-Nearest Neighbors (KNN), Gradient Boosting (GB), Logistic Regression (LR), Decision Tree (DT), and Random Forest (RF) to classify 11 dynamic hover hand gestures. Our findings indicated that RF achieved the highest accuracy among all other ML classifiers, with more than 97% accuracy. The implications of our study could inform researchers of the potential of this technology for recognising accurate hand gestures trained with multiple subjects, which could help them establish gesture recognition systems for improved accuracy and usability.
