https://bura.brunel.ac.uk/handle/2438/8623 Tue, 03 Nov 2026 02:07:04 GMT 2026-11-03T02:07:04Z https://bura.brunel.ac.uk/handle/2438/34071 Title: Dynamic data reduction for foundational AI models Authors: Sanderson, Dominic Abstract: Arti cial Intelligence (AI) fundamentally relies on data to generate useful outputs. The accuracy of an AI system is not only determined by the e ectiveness of the model, but also both the quantity and quality of the data it is trained on. Therefore, a relatively simple way to maximise AI performance is to use more data. However, as we continue to push the boundaries of AI, the demand for larger and more diverse datasets has only intensi ed, highlighting the importance of high quality data acquisition and management in the development of advanced AI technologies. This causes a growing concern about the environmental impact associated with these methods. The escalating energy requirements for training and operating sophisticated AI models contribute to a larger carbon footprint, prompting discussions on sustainable practices within the eld of arti cial intelligence. Addressing these concerns necessitates a balance between pursuing technological advancements and implementing strategies to mitigate their environmental repercussions. This thesis explores dynamic data reduction methods as a means of exploring data quality and usefulness. For example, by reducing the amount of training data used, less computation is needed, reducing the environmental cost of the algorithm. Random dynamic data reduction is explored rst by reducing the amount of data used for training at di erent stages of training, for image classi cation tasks. Three dynamic data reduction methods are introduced to observe di erent strategies of data reduction: data step, data increment, and data cut. These strategies present varying amounts of training data application which means that fewer image evaluations are performed to train the model than if the model is trained on the full dataset. Meanwhile, the strategies ensure that the full dataset is used, to ensure maximum data coverage. Results show that the data step method is best performing, and for datasets MNIST and smallNORB, model accuracy may be increased by 0.04% and 0.593% respectively, while 12% and 6% fewer evaluations image are performed. Random data reduction is so far limited by uncertainty of the quality of data that is removed for a portion of the training. This may be alleviated by introducing a method to determine speci c data to remove. By rst applying dimensionality reduction algorithms, data may be ordered and then removed based on distance from each classes' centroid. This allows for more intelligent data selection. Data step reduction is performed on larger image datasets, namely Imagenette and the Micro-PCB dataset, yielding inferior accuracy for the Imagenette dataset with an accuracy decrease of 0.114%, and superior accuracy for the Micro-PCB dataset by 0.013%. The data step method has shown promising results as a dynamic data reduction strategy. An alternative method using a sliding window of data is explored, which ensures unique data each epoch throughout training, instead of a step-up in data use. This method is combined with the previously used non-random data reduction, on image data of di ering sizes. Results show that model accuracy is worse for MNIST, CIFAR-10, Imagenette and the Micro-PCB datasets by up to 0.16%, but accuracy increases for the smallNORB dataset by 0.654%. However, statistical signi cance against benchmark experimentation informs us that after reducing data, results show no statistically signi cant di erence from the benchmark, even though training data has been reduced. Still, despite the underwhelming performance, insight is given into which data is more useful at the beginning or end stages of training by comparing the direction of the sliding data window. Results show that for unstructured datasets MNIST, CIFAR-10 and Imagenette, the forward data ow direction yields superior accuracies, while for the highly structured smallNORB and Micro-PCB datasets the opposite is true. This supports the theory that some data is more useful in the ne-tuning stages of training, depending on how structured the data is. Having explored dynamic data reduction in image data, data reduction in imbalanced, timeseries occupancy datasets is investigated. Occupancy detection from non-invasive sensor readings is useful for building interior management, while maintaining personal privacy. However, realworld data is often imbalanced, often giving bias to training. By using data reduction techniques inspired by those used for image data, results show that data reduction on the larger class only can maintain and even improve AUC-ROC scores 0.483% with as much as 39% of the dataset removed. Dataset fusion is also explored as a means to improve the generalisability of the classi er. Results of experimentation show a decrease in AUC-ROC of as much as 48% with the full dataset, and 33% with the reduced dataset. This shows that by attempting to increase the available data, overall performance is decreased as the model struggles to specialise. Also, data reduction on a fused dataset can minimise the detrimental e ects of too much data. Class density is also calculated to be a metric that measures class balance and data su ciency, which is vital for determining whether a dataset may be reduced to improve classi cation performance prior to any model training. The above investigations show that the exclusion of some data may improve model performance, as well as reduce training cost. The novel data reduction strategies that are introduced also give insight into which data is more important to training at the earlier and later stages. These ndings prompt us to consider not just the quantity of data, but the quality and relevance of data during both collection, where focus can be aimed at collecting most useful data, and training, where most relevant data may be applied at the most bene cial stage of training. Description: This thesis was submitted for the award of Doctor of Philosophy and was awarded by Brunel University London Thu, 01 Jan 2026 00:00:00 GMT https://bura.brunel.ac.uk/handle/2438/34071 2026-01-01T00:00:00Z https://bura.brunel.ac.uk/handle/2438/33761 Title: Innovations in arithmetic for neuromorphic computing Authors: Fernandez-Hart, Tim Abstract: The continued growth in computational demand, driven in large part by machine learning workloads, has coincided with the slowing of performance gains from semiconductor scaling alone. Therefore, additional computing power can no longer be simply obtained by using increased transistor density or clock frequency to scale standard architecture. Instead, alternative computing paradigms are required. Neuromorphic computing has emerged as one such approach, offering architectures designed to improve computational efficiency by mimicking aspects of biological neural networks such as massive parallelism, locality of memory and computation, and sparse activity. Spiking Neural Networks (SNNs) are the primary computational models executed on neuromorphic processors, processing information through temporal dynamics and sparse events rather than dense numerical operations. However, the numerical representation used to implement these dynamics remains an open design question. This thesis investigates the role of arithmetic choice in digital neuromorphic computing, with a particular focus on posit arithmetic as an alternative to conventional fixed-point and IEEE-754 floating-point formats. Rather than treating reduced precision as an optimisation problem to be mitigated against, this work adopts an arithmetic-centric perspective, evaluating whether alternative number systems can intrinsically offer improved numerical behaviour at reduced bit-widths. To this end, rather than the more common mixed-precision approach, all experiments use reduced-precision emulation throughout to simulate a neuromorphic system operating natively with the target format, providing a conservative assessment of arithmetic suitability for future neuromorphic hardware. The thesis makes three primary contributions. First, it presents a systematic study of reduced-precision arithmetic for simulating neural dynamics, using the Izhikevich neuron model across its full range of canonical firing patterns. The results demonstrate that 16-bit posit arithmetic can match or outperform IEEE-754 floating-point equivalents of equal bit-width, and in some cases 16-bit posit can match 64-bit floating-point accuracy, if combined with a simple rescaling strategy. Second, the work provides the first comprehensive evaluation of 8-bit posit arithmetic for offline SNN training using gradient-based methods. Across both frame-based and event-based datasets, 8-bit posits are shown to enable effective training without the need for mixed-precision schemes, loss scaling, or other mitigation techniques commonly required by low-precision floating-point formats. Finally, the thesis presents the first detailed analysis of arithmetic effects on online learning in SNNs using the e-prop algorithm, demonstrating that 16-bit posits enable stable training and outperform higher-precision floating-point formats, while fixed-point representations require approximately 64-bit precision to attain similar performance. Collectively, these results show that posit arithmetic can match the numerical and system level performance of larger floating point formats while using smaller bit widths and requiring fewer, if any, mitigation strategies. The designers of future digital neuromorphic architectures should therefore seriously consider including posit rather than fixed or floating point arithmetic units. Description: This thesis was submitted for the award of Doctor of Philosophy and was awarded by Brunel University London Thu, 01 Jan 2026 00:00:00 GMT https://bura.brunel.ac.uk/handle/2438/33761 2026-01-01T00:00:00Z https://bura.brunel.ac.uk/handle/2438/33450 Title: AI-enabled flaw detection using multi-sensory data fusion Authors: Marsh, Benedict Abstract: This thesis focuses on the challenge of automated flaw detection by developing a method for surface crack detection that uses a data fusion approach. Flaw detection is needed to detect damage that could compromise structural integrity, leading to further consequences. Identification of flaws is important to analyse the severity and then take action to rectify any issues. Automated approaches using AI are needed to reduce cost as well as to speed up identification and increase accuracy. The presented research developed a method that followed a multi-stage approach, where data from multiple sensors are fused into a 3D representation with the use of AI models. Then, detection is done on that representation to identify the cracks so that further analysis can be done to determine the crack severity. Research contributions are from both stages. First, data fusion improvements for RGB images were worked on, and a novel method for fusing depth data from RGB stereo and LiDAR data was developed. Then, a method for crack identification from RGB-D data using a novel synthetic data generation method was developed. Evaluation of the contributions was carried out to demonstrate the improvements due to data fusion with comparisons to other methods. The key findings included: metric evaluations showing the developed enhanced image fusion technique improved RGB image quality. The developed novel stereo and LiDAR data fusion method showed lower error than either input method alone. The evaluation of the developed novel crack segmentation method using synthetic data showed that models can be effectively trained in the absence of extensive real-world data. Description: This thesis was submitted for the award of Doctor of Philosophy and was awarded by Brunel University London Wed, 01 Jan 2025 00:00:00 GMT https://bura.brunel.ac.uk/handle/2438/33450 2025-01-01T00:00:00Z https://bura.brunel.ac.uk/handle/2438/33449 Title: The multi-sensor fusion of camera and LiDAR with semantic 3D depth sensing for enhanced perception in autonomous driving systems Authors: Yildiz, Ahmet Serhat Abstract: The development of autonomous driving technology depends on the precise perception of the environment and effective object detection and classification, tracking, essential for safe navigation and decision-making. However, existing multi-sensor fusion methods focus on generating dense depth maps by combining camera and LiDAR data, where distance is represented using colour values. In these depth maps, object shapes are often unclear and small or partially occluded objects are difficult to detect. As a result, it is challenging to accurately detect an object and extract object’s specific distance information. Therefore, this research focuses on detecting specific important objects and directly extracting their distance information from a sparse depth map combined in a more efficient and reliable way. This thesis explores multi-sensor fusion strategies, specifically fusing camera and LiDAR data, to enhance the perception capabilities of autonomous driving systems. While LiDAR sensors provide accurate depth measurements, their limited resolution restricts detailed scene understanding. Conversely, cameras provide much semantic information, but lack depth precision. This research addresses sensor limitations by integrating YOLOv8, a state-of-the-art object detection framework, with LiDAR point cloud data using precise camera–LiDAR projection utilising calibrated transformation matrices. The system is evaluated using the KITTI object detection benchmark, demonstrating improved range resolution and detection robustness under complex driving conditions. This work presents three main contributions: CNN-based traffic sign classification, camera-LiDAR fusion with depth enhancement and projection, and novel depth estimation techniques for distance measurement. Firstly, a Convolutional Neural Network (CNN) model is developed, trained, and implemented on the German Traffic Sign Recognition Benchmark (GTSRB) to achieve reliable classification of traffic signs under varying real-world conditions. The CNN architecture has several convolutional layers with activation functions (including ReLU, Leaky ReLU, and GELU), and each one is followed by a max-pooling layer that systematically reduces the spatial dimensions while keeping important features. After that, the extracted features are subsequently sent to fully connected (FC) layers for classification purposes. The network utilises the Adam optimiser with categorical cross-entropy loss and employs regularisation methods. The study evaluates and compares the performance of different activation functions to analyse their impact on recognition accuracy and model robustness. The final model does a great job of classifying traffic signs in the GTSRB test set, offering a dependable vision input source for further perception tasks in autonomous driving systems. Secondly, this study introduces a complete camera–LiDAR fusion framework that improves depth perception by using calibration data and transformation matrices, which include both intrinsic and extrinsic parameters, to project 3D LiDAR point clouds onto 2D camera images. Using homogeneous coordinate transformations, and matrix multiplication, the projection pipeline carefully maps LiDAR points onto the image plane, creating a sparse depth map that matches the RGB data. LiDAR sensors are naturally sparse and have low resolution, especially in systems with fewer vertical beams. To address this problem, the interpolation method was used to make the depth map denser and emulate higher-resolution point distributions. This upsampling process was used on both bounding box and segmentation mask regions to evaluate the efficacy of various spatial priors. The study provides a foundation for future perception pipelines in autonomous driving systems. Thirdly, a novel distance estimation method is proposed based on fused cam-era–LiDAR data. Several depth extraction techniques are introduced and evaluated, including Point-by-Point (PbyP), Complete Region Depth Extraction (CoRDE), Central Region Depth Extraction (CeRDE), and Grid Central Region Depth Ex-traction (GCRDE). These methods are tested across various object categories (e.g., cars, trucks, bicycles) and occlusion levels (0 to 3) using metrics such as extraction time, accuracy, and Root Mean Square Error (RMSE). Results show that segmentation mask-based methods, especially CeRDE and GCRDE, achieve higher depth estimation accuracy and lower RMSE, particularly for large and occluded objects. However, bounding box methods like PbyP and CoRDE maintain faster processing times, favoring real-time applications. GeRDE provides a balanced solution, offering both high accuracy and computational efficiency. Overall, this thesis contributes to the field of autonomous driving systems perception by demonstrating that deep learning-enhanced sensor fusion and optimised depth extraction can significantly improve the performance and reliability of perception systems under complex real-world conditions. Description: This thesis was submitted for the award of Doctor of Philosophy and was awarded by Brunel University London Thu, 01 Jan 2026 00:00:00 GMT https://bura.brunel.ac.uk/handle/2438/33449 2026-01-01T00:00:00Z