Abstract Siamese Neural Networks (SNNs) have shown promise in addressing a variety of tasks, even with limited data availability. However, their adoption is hindered by the lack of transparency in their decision-making processes. A key challenge in explaining SNNs lies in the absence of an inverse mapping between high-dimensional input feature vectors and the low-dimensional embedding space. Therefore, computing direct distances between input features becomes meaningless. Existing autoencoder-based explanation methods face several limitations. These include poor image reconstruction quality due to insufficient data and the omission of final distance layer of the SNN during the explanation process. While the Siamese Network Explainer (SINEX) can explain audio and grayscale images, it does not support RGB images. To overcome these challenges, we propose a method called Features Distance-based eXplanation (FDbX). This approach identifies salient features using ridge regression, trained on perturbed SLIC-segmented images. To enhance the selection of important features, we incorporate Bayesian analysis, which assigns importance scores to features. To provide a comprehensive explanation of the decision route, we construct a mathematical model that represents important features and their Hamming distances as a bipartite graph. In this graph, nodes represent features and edges denote distances between feature pairs. The resulting explanation heatmaps highlight critical image segments, offering more intuitive and visually informative explanations than existing methods. We evaluate stability and faithfulness of our method using stability indices such as $$R^2$$ and mean squared error. To the best of our knowledge, this is the first work to introduce Variable and Coefficient Stability Indices for image datasets.

Malicious Software (Malware) is a primary element of many cyber crimes and attacks, causing massive damage and financial losses to organizations. Accordingly, malware detection and classification has become a crucial security field resulting in various attempts from researchers to develop solutions including signature-based approaches to Artificial Intelligence (AI) models showing their efficacy in detecting malware. Yet, users still have reservations about AI models due to the ambiguity and mysteriousness of their decisions resulting from their black-box nature. To address this problem, this paper proposes to develop explainable AI models to classify the malware families robustly. The proposed method uses text classification of API call sequences generated by these families by considering two datasets. A weighted training methodology is used to solve the dataset imbalance problem. Subsequently, the method presents an eXplainable AI (XAI) approach to establish an understandable and interpretable relationship between the API call sequences and the Bidirectional Encoder Representations from Transformers (BERT) model decisions, which enhance the model accountability and usability by employing the Local Interpretable Model-Agnostic Explanation (LIME) and the Shapley Additive Explanations (SHAP) platforms. The results reveal that the BERT model outperforms its counterparts considering F1 score, Balanced Accuracy (BA), and Matthews correlation coefficient (MCC).

Building upon pre-trained ViT models, many advanced methods have achieved significant success in COVID-19 classification. Many scholars pursue better performance by increasing model complexity and parameters. While these methods can enhance performance, they also require extensive computational resources and extended training times. Additionally, the persistent challenge of overfitting, due to limited COVID-19 dataset sizes, remains a hurdle. To address these challenges, we proposed a novel method to optimize pre-trained transformer models for efficient COVID-19 classification with stochastic configuration networks (SCNs), referred to as OPT-CO. We proposed two optimization methods: sequential optimization (SeOp) and parallel optimization (PaOp), by incorporating optimizers in a sequential and parallel manner, respectively. Our method can enhance model performance without necessitating a significant parameter expansion. Additionally, we introduced OPT-CO-SCN to avoid overfitting problems through the adoption of random projection for head augmentation. The experiments were carried out to evaluate the performance of our proposed model based on two publicly available datasets. Based on the evaluation results, our method achieved superior, performance surpassing other state-of-the-art methods.

Abstract: Recent advances in VR/AR have fueled the interest in automating the development of 3D interfaces. An important step of this automation is 3D object recognition. This allows applications to pre-configure assets based on expected behaviours, rather than requiring human developers to hard code such assets. This paper presents a clustering approach for reliably identifying 3D objects. Model-based descriptors are used to extract key features and geometric data are used to develop models for 3D object recognition. Two clustering approaches are used in this paper: hierarchical and K-means. The models are tested on two industrial gearboxes and an accuracy score is calculated to test model performance. The two gearboxes were too small resulting in the accuracy of both algorithms being identical, however the computational efficiency of the K-means algorithm makes it a more suitable choice. The accuracy for both algorithms was 74% when tested on a combined dataset of both gearboxes. This increased to 81% when difficult to identify parts are removed. The model can reliably identify geometrically consistent parts like ball bearings, roller bearings and screws.

With the Internet of Things technology, almost any remote sensing devices, wearables, and smart objects are equipped to transmit large volumes of data in continuous streams. In conventional cloud-centric analytics, all the raw data is transferred to a data centre and processed in real-time, near real-time, or in batches. However, this approach is usually not very responsive to real-time analytics due to the latency in transmission alongside network traffic, bandwidth and data transmission costs. To tackle this, edge-enhanced analytics ensures that raw data can be preprocessed at the edge and sent across the network channel in a more compact form. A specific category of deep learning model, autoencoder, can help to achieve this by transforming high-dimensional data into compact representation. We propose an edge-enhanced framework which deploys a deep autoencoder model on the network edge for data compression. After training of models in the cloud, the encoder part of the autoencoder is deployed on the edge for data reduction while the decoder remains on the cloud to reconstruct the data for an image classification task on the cloud. We applied supervised fine-tuning using the intrinsic dimensionality of the data to achieve an accuracy that surpasses the baseline cloud model. The solution was explored in the context of an image recognition problem using the MNIST and FASHION-MNIST datasets. The framework was validated on an event simulator to estimate the network savings of the proposed method in terms of bandwidth and latency. The edge-enhanced approach saves up to 74% bandwidth compared to the centralised analytics. In addition, real-time analytics is further improved by taking 25% less time to complete the task.

The diversity and proliferation of Knowledge bases have made data integration one of the key challenges in the data science domain. The imperfect representations of entities, particularly in graphs, add additional challenges in data integration. Graph dependencies (GDs) were investigated in existing studies for the integration and maintenance of data quality on graphs. However, the majority of graphs contain plenty of duplicates with high diversity. Consequently, the existence of dependencies over these graphs becomes highly uncertain. In this paper, we proposed graph probabilistic dependencies (GPDs) to address the issue of uncertainty over these large-scale graphs with a novel class of dependencies for graphs. GPDs can provide a probabilistic explanation for dealing with uncertainty while discovering dependencies over graphs. Furthermore, a case study is provided to verify the correctness of the data integration process based on GPDs. Preliminary results demonstrated the effectiveness of GPDs in terms of reducing redundancies and inconsistencies over the benchmark datasets.

Deep Reinforcement Learning (DRL) has been proven to be capable of designing an optimal control theory by minimising the error in dynamic systems. However, in many of the real-world operations, the exact behaviour of the environment is unknown. In such environments, random changes cause the system to reach different states for the same action. Hence, application of DRL for unpredictable environments is difficult as the states of the world cannot be known for non-stationary transition and reward functions.

Video analytics systems based on deep learning approaches are becoming the basis of many widespread applications including smart cities to aid people and traffic monitoring. These systems necessitate massive amounts of labeled data and training time to perform fine tuning of hyper-parameters for object classification. We propose a cloud based video analytics system built upon an optimally tuned deep learning model to classify objects from video streams. The tuning of the hyper-parameters including learning rate, momentum, activation function and optimization algorithm is optimized through a mathematical model for efficient analysis of video streams. The system is capable of enhancing its own training data by performing transformations including rotation, flip and skew on the input dataset making it more robust and self-adaptive. The use of in-memory distributed training mechanism rapidly incorporates large number of distinguishing features from the training dataset - enabling the system to perform object classification with least human assistance and external support. The validation of the system is performed by means of an object classification case-study using a dataset of 100GB in size comprising of 88,432 video frames on an 8 node cloud. The extensive experimentation reveals an accuracy and precision of 0.97 and 0.96 respectively after a training of 6.8 hours. The system is scalable, robust to classification errors and can be customized for any real-life situation.

The recent rise in multimedia technology has made it easier to perform a number of tasks. One of these tasks is monitoring where cheap cameras are producing large amount of video data. This video data is then processed for object classification to extract useful information. However, the video data obtained by these cheap cameras is often of low quality and results in blur video content. Moreover, various illumination effects caused by lightning conditions also degrade the video quality. These effects present severe challenges for object classification. We present a cloud-based blur and illumination invariant approach for object classification from images and video data. The bi-dimensional empirical mode decomposition (BEMD) has been adopted to decompose a video frame into intrinsic mode functions (IMFs). These IMFs further undergo to first order Reisz transform to generate monogenic video frames. The analysis of each IMF has been carried out by observing its local properties (amplitude, phase and orientation) generated from each monogenic video frame. We propose a stack based hierarchy of local pattern features generated from the amplitudes of each IMF which results in blur and illumination invariant object classification. The extensive experimentation on video streams as well as publically available image datasets reveals that our system achieves high accuracy from 0.97 to 0.91 for increasing Gaussian blur ranging from 0.5 to 5 and outperforms state of the art techniques under uncontrolled conditions. The system also proved to be scalable with high through-put when tested on a number of video streams using cloud infrastructure.

Evidence-based medicine is critically dependent on three sources of information: a medical knowledge base, the patient's medical record and knowledge of available resources, including where appropriate, clinical protocols. Patient data is often scattered in a variety of databases and may, in a distributed model, be held across several disparate repositories. Consequently addressing the needs of an evidence- based medicine community presents issues of biomedical data integration, clinical interpretation and knowledge management. This paper outlines how the Health-e-Child project has approached the challenge of requirements specification for (bio-) medical data integration, from the level of cellular data, through disease to that of patient and population. The approach is illuminated through the requirements elicitation and analysis of Juvenile Idiopathic Arthritis (JIA), one of three diseases being studied in the EC-funded Health- e-Child project.