Talk by Negar Safinianaini. (The Royal Institute of Technology)

Speaker: Negar Safinianaini (The Royal Institute of Technology)

Title: The past and future research on Bayesian-machine-learning tools to bridge gaps between life science and current machine learning tools

Abstract: It is natural to model the physical world in which the data is formed; the mathematical modeling translates the physical events, expressed as input data, to mathematical characteristics, which are helpful in pattern recognition computations. This approach is commonly referred to as the model-based approach, followed by a so-called Bayesian machine learning tool that performs inference. Bayesian-machine-learning tools are favored for the other machine learning approaches, such as feature-based and deep learning, when it comes to providing measures of uncertainty, allowing for rapid prototyping, allowing for interpretability, and enabling a probabilistic framework. This talk, firstly, covers the past research in which novel Expectation-Maximization (EM) and Variational Inference (VI), instances of Bayesian-machine-learning tools, are devised to bridge gaps between life science and machine learning. Secondly, the talk proposes two intermingled research directions to provide adaptive Bayesian-machine-learning tools. The first direction is to expand VI to allow for dynamical, adaptive, and context-aware learning. The second direction is to improve cancer analysis by mapping informative parts of biological data to the model, followed by novel EM and VI tools. The second direction is a free sandbox for the first direction, and it contributes to detecting and concretizing unknown issues and statistical behaviors in VI and EM, as well as providing solutions to those issues.