Syvät konvolutionaaliset gaussiset prosessit
Kenneth Blomqvist · Aaltodoc (Aalto University) · 2019
Convolutional neural networks have achieved unparalleled results on various machine learning tasks such as image classification, speech recognition, image segmentation, machine translation and many others. Modern neural network architectures have millions of parameters. This makes them prone to overfitting and sensitive to out-of-sample noise. As they are relatively practical to train, these issues can often be counteracted using massive amounts of training data. They have also been found to be prone to adversarial attacks. Developing methods which are well-regularized and could learn complicated functions without using massive amounts of data could enable us to deploy machine learning methods in settings where heaps of data are not available. Gaussian processes are known as a well-regularized statistical method which works beautifully for simple regression and classification tasks with a small number of training examples. Achieving such properties in deep models would be greatly beneficial. In this thesis we develop a deep Gaussian process model with convolutional structure which we call the deep convolutional Gaussian process. It is a method for modelling hierarchical combination of local features using Gaussian process mappings structured in a hierarchical manner. We compare our method on the MNIST and CIFAR-10 image classification tasks against other successful approaches. On the CIFAR-10 dataset, we achieve a more than 10\% improvement in test classification accuracy over other Gaussian process based methods.