| dc.description.abstract | Nowadays, machine learning and deep learning are present in the most diverse types of applications. From such diversity, many particular designs, architectures, training methods were created, given the variety of applications of many different areas. Rather than trying to find a multi-purpose scheme to generate and train deep neural networks, which would be an impracticable challenge, this work aims to deliver well suited training techniques for well defined problems in specific fields. Every single domain has its own necessities and specificities, thus it is essential to have individual solutions for each case.
In the bio-informatics domain we propose the gGAN, a novel approach to train GANs, which is capable of generating labeled genetic datasets using a small labeled dataset and a larger unlabeled dataset, exploiting concepts of semi-supervised learning and data augmentation to create a new approach to deal with the limited labeled data available to researchers. This method may also be used as a self-aware classifier, a classifier with a second level of confidence. Since it is only based on genetic profiles, it can be applied at any stage of the disease (or even before infection). This allows the usage as a triage tool, able to prognose early-infected patients, avoiding the exposure of healthcare professionals who are sensitive to the disease.
This work also addresses a Scientific Machine Learning technique, the Physically Informed Neural Networks (PINNs). Evidence shows that PINN training by gradient descent displays pathologies that often prevent convergence when solving PDEs with irregular solutions. In this work, we propose the use of a Particle Swarm Optimization (PSO) approach to train PINNs. The resulting PSO-PINN algorithm not only mitigates the undesired behaviors of PINNs trained with standard gradient descent but also presents an ensemble approach to PINN that affords the possibility of robust predictions with quantified uncertainty. Comprehensive experimental results show that PSO-PINN, using a modified PSO algorithm with a behavioral coefficient schedule, outperforms other PSO variants for training PINNs, as well as PINN ensembles trained with standard ADAM.
Furthermore, we propose two distinct extensions for this method, namely Multi-Objective PSO-PINN and Multi-Modal PSO-PINN. The first one acknowledges the PINN as a multi-objective problem and handles the PSO-PINN training as such. This approach unleashes a new paradigm to deal with PINNs, allowing the analysis of the problem and finding out if the model-driven and data-driven portions of the PINN are in agreement. The latter promotes a desirable characteristic of the PSO-PINN, the diversity of the solutions. The Multi-Modal approach enforces multiple local optima during the training, guaranteeing a solution composed of a diverse ensemble. | |
