A0753
Title: Mixture of experts models for statistical learning in the ocean
Authors: Jacob Bien - University of Southern California (United States) [presenting]
Sangwon Hyun - University of California, Santa Cruz (United States)
Oh-Ran Kwon - BioInfra Inc (Korea, South)
Francois Ribalet - University of Washington (United States)
Mattias Cape - University of Washington (United States)
Abstract: The mixture of experts (MoE) model provides a flexible framework for modeling complex phenomena. Two case studies are described in which data collected by oceanographers can be conveniently modeled as an MoE. Case 1: Although microscopic, phytoplankton in the ocean are extremely important to all of life and are together responsible for as much photosynthesis as all plants on land combined. Oceanographers measure the distribution of different types of phytoplankton through flow cytometry conducted onboard a moving ship. A novel sparse multivariate MoE model is presented to predict the time-varying phytoplankton subtypes based on environmental covariates. In this application, each expert corresponds to a distinct subtype of phytoplankton, and the gating function captures the relative abundances of the different subtypes. Case 2: Direct measurement of chlorophyll concentration by oceanographers is time-intensive, and therefore, oceanographers rely on satellite data, which is abundant, global, and up-to-date. Algorithms estimate chlorophyll concentration based on satellite-measured ocean color. It is shown that a state-of-the-art algorithm used for this prediction task can be interpreted as an MoE model. Interestingly, the learning approach is semi-supervised, leveraging the large amount of satellite data available. A variety of MoE models are explored in the context of this important prediction problem.
