Bayesian Learning with Unbounded Capacity from Heterogeneous and Set-Valued Data
Abstract
The aim of this project was to advance machine learning methods grounded in Bayesian theory, optimal transport, point processes and random finite sets to deal with growing complexity and heterogeneity of large-scale data. The research team has focused on two main themes: i) developing necessary theory to perform parameter estimation with latent variables for set-valued data using point process theory, and ii) formulating and developing fast inference procedures for Bayesian models via Wasserstein and optimal transport theory. Both of these themes are related through their roles as the building blocks to construct more advanced and scalable models to deal with not only standard data types (such as vectors and matrices), but also set-valued data. The most important results are twofold: a) new model-based method to reason and learn from set-valued (aka point pattern data).This includes new models for classification, clustering and novelty detection with point pattern data; and its extension to deal with sequential data. And, b) a new Wasserstein-based formulation for multi-level clustering in high-dimensional data; this formulation also provides a new scalable framework for Bayesian inference and as opposed to traditional information-theoretic approaches. These results have been documented in 6 research papers where four have been accepted for publication (DSAA, ICPR and ICML),one is under review (Pattern Recognition) and another one is under preparation to be submitted to Journal of Machine Learning Research.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 13, 2018
- Accession Number
- AD1057702
Entities
People
- Bao-tu Ho
- Dinh Phung
Organizations
- Deakin University