Learning to Build Probabilistic Models with Limited Data
Jake Snell · TSpace (University of Toronto) · 2021
Deep learning has successfully transformed a wide range of machine learning applications in recent years. One of its keys to success is the ability to learn relevant features from scratch on large amounts of data. Yet there are many real-world scenarios in which a new model must be built quickly with only a small amount of data. This naturally raises the question: How can we use deep learning to efficiently build task-specific models given limited data? In this thesis, we introduce a framework for understanding relationships among these sorts of problems and propose several novel algorithms for tackling challenging tasks in this regime including few-shot learning, zero-shot learning, and structured prediction. In the first part of this thesis, we propose a deep metric learning approach to few-shot and zero-shot classification that generalizes from just a few examples. It does so by exploiting a simple inductive bias, namely that there exists an embedding space in which examples from the same class are clustered around a single prototype representation. We demonstrate the strong accuracy of our approach on standard few-shot and zero-shot benchmarks. We then consider image segmentation, an task in which is important for a machine learning system to capture multiple plausible outputs due to ambiguity. We introduce a tree-structured graphical model for representing such distributions and propose an algorithm that optimizes the model structure for each image. When evaluated on a challenging image segmentation datasets, our algorithm is successfully able to generate multiple diverse segmentations much in the way that humans do. Finally, we develop a Bayesian approach to few-shot classification based on Gaussian processes. Our method learns a deep kernel to compute the similarities among input images and uses a Pólya-Gamma augmentation scheme to achieve tractable inference. We show strong predictive accuracy on several challenging few-shot learning datasets and improved uncertainty quantification over baseline methods.