Deep Learning and Neural Networks (DNNs) development has seen an unprecedented expansion during the last couple of years, as it allows scientists from multidisciplinary fields to boost their domain specific discoveries with the computational power of DNNs. One of the key aspects of the DNNs is the ability to investigate the conventional multi-perceptron deep neural structure as a hierarchy (e.g., input, decisional or output units), in a manner that suits the real-world study cases, their inner structuration, calibration and limits. The improved neural structure could be tailored to the domain specifications, more efficient and certified to be trustworthy (e.g., graph connected features for time-variant phenomena analysis, a neat, transparent inference for handwritten images, etc). During my research journey, I was able to dig into the “black box” DNN structure, motivate the restructuration of the input unit, in particular, to embed features (nodes and edges embedding) into a reduced optimal vector. Consequently, a transparent and an explainable graph-based solution for the input data space processing. Moreover, I further examined the decisional unit of DNNs, and improved both the forward and backward learning paths with a neat feature selection. As opposite to conventional DNNs, I alleviated the matricial complexity of the conventional input sampling and the gaussian measure for features' selection with a structural connectionist metric of the maximum connectivity for each feature. This structural mapping allows a scored backpropagation process and a total average CPU time reduction of ~36.4% with ~83% accuracy. I improved the previous connectionist mapping with a deterministic graph traversal ability. The feature importance criterion at this stage is determined through the betweenness centrality around the nodes expressing time variance. The random graph search was improved by a provable and a functional acyclic graph property that helps decorate the necessary features during the backpropagation process and achieve ~70% less feature calibration with ~99.68% accuracy. I bridged the input data sampling strategy with the neural activations using the observability Gramian global maxima. Consequently, the number of neurons activating features will be determined while achieving the optimal accuracy of each layer. Only ~13.6% of necessary paths were activated achieving ~99.9% accuracy. I achieved a convex, and a reduced calibration of the random, unexplained neural activations where only ~11% of the input space constituents deemed necessary for training. Subsequently, I analysed the neural topology and enriched the conventional bias, weight thresholds with a centrality measure that serves as an automatic filter to decide the necessary number of neurons, features, for each hidden or activated neural layers, while preserving an optimal model performance. I tested our developed models on multi-structured public datasets (more than 12 Health news datasets issued from Twitter API, Covid-19 tweets and covid variants, MNIST handwritten digit images and EMNIST for handwritten letter images). The obtained results outperformed state-of-the-art performance (e.g., ~83% accuracy with Health tweets classification, 99.86% accuracy with Covid-19 tweets sentiment prediction joined with Covid variants analysis, 99.8% accuracy with MINST, EMNIST image classification.). The developed input characterization along with the formalized feature selection was proved to be certified, generalizable, through a formal graph properties verification. Conclusions that derive a DNN calibration with less, necessary parameters, regardless the waste of research effort in calibrating the fully connected, statistical conventional DNN versions, and may suggest a wider application within artificial neural development.
| Date of Award | 15 May 2025 |
|---|
| Original language | English |
|---|
| Supervisor | Richard Hill (Main Supervisor) & Qiang Xu (Co-Supervisor) |
|---|