Machine Learning Methods for Autonomous Classification and Decision Making

This thesis focuses on developing machine learning methods for autonomous classification and decision making, especially on two case studies: traffic speed prediction and cancer bone segmentation. For traffic speed prediction, the convolutional neural network (CNN) achieves state-of-the-art results in complex traffic networks. However, the pooling layers cause the loss of information within the data. This thesis proposes an efficient capsule network for traffic speed prediction. The proposed capsule network replaces the pooling layer with capsules connected by dynamic routing and encodes the features and probability of those features showing on the local region. The proposed capsule network provides outperformed results compared to state-of-the-art CNNs. However, the CNN and capsule network (CapsNet) are parametric models and the uncertainty is, thus, not analysed. Two Gaussian process (GP) frameworks are proposed for traffic speed prediction, equipping the CNN with the ability to quantify uncertainty. The first framework proposes to equate a state-of-the-art CNN with a shallow GP. The proposed approach is evaluated and the uncertainty is analysed by applying the confidence interval. In addition, the impact of the noise is investigated by adding a different level of noise. The second framework is a novel deep kernel CNN-GP framework with spatio-temporal kernels, allowing it to abstract high-level features and consider both time and space. The proposed CNN-GP framework is validated and evaluated using CO2 concentration and traffic prediction for the short-term and long-term. An efficient uniform error bound is proposed and evaluated with simulated and real data. For cancer bone segmentation, machine learning methods are proposed to segment bone lesions in cancer-induced bone disease from Micro Computed Tomography (µCT) images, which brings a new perspective of dealing with bone caner segmentation. The performances are evaluated and their effectiveness is compared. Due to the limited number of datasets and the lack of labelled lesions within the dataset, an approach to generate simulated data is proposed. With an enhanced dataset, a generative adversarial network is proposed to reconstruct the bone with a lesion to a healthy bone. Consequently, the location of the lesion can be obtained by subtracting the original image from the reconstructed image.