Recurrent and Bayesian Kernel Learning for Small Data with Applications to Neuroimaging

• Author / Creator

  Lewandowski, Alex

• Gaussian processes are flexible probabilistic models for regression and classification.
  However, their success hinges on a well-specified kernel that can
  capture the structure of data. For complex data, the task of hand crafting a
  kernel becomes daunting. In this thesis, we propose new methods for Gaussian
  process classification. In particular, we propose learning flexible kernels
  that are parameterized by recurrent neural networks. Unlike previous work
  in recurrent kernel learning, our recurrent kernels are learned through a scalable
  variational framework and applied to classification. We also investigate
  methods to propagate the uncertainty of neural network parameters through
  the Gaussian process. In doing so, we propose a novel use of batch normalization
  that provides estimates of uncertainty for feed-forward and convolutional
  neural networks. To evaluate our models, we simulate multivariate time-series
  data with commonly observed phenomenon in neuroimaging data and compare
  against two baselines: Gaussian processes and neural networks. We then
  compare our model to the best performing baselines on common deep learning
  benchmarks and real neuroimaging datasets. We find that our models provide
  better predictions and estimates of uncertainty when sample sizes are small
  and remains competitive at larger sample sizes.

• Subjects / Keywords

  • Bayesian deep learning
  • Kernel learning
  • Gaussian process classification
  • Machine learning
  • Neuroimaging

• Graduation date

  Fall 2018

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/R3RR1Q353

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.