Mixtures of Neural Network Experts with an Application to Phytoplankton Flow Cytometry Data

Music City Center 

Analysis of flow cytometry data allows oceanographers to identify and distinguish between different types of photosynthetic microbes, called phytoplankton. Recent development of flow cytometry data analysis has included a gradual increase in the use of model-based clustering, the utility of which depends upon high clustering accuracy as well as cogent interpretations of the relationships between cells' optical properties and environmental conditions, such as sunlight intensity, temperature, salinity, and nutrient concentrations. Here, we improve the latter aspect via a mixture of experts which utilizes random weight neural networks, thereby flexibly estimating the dependence of cell types' optical properties and relative abundances upon environmental covariates without the computational burden of training by backpropagation. We show that the proposed model provides better out-of-sample pointwise predictive accuracy and more realistic interpretations of phytoplankton behaviors than mixtures of linear experts in a variety of simulated scenarios and an application to real flow cytometry data.

flow cytometry

mixture of experts

neural network

random weight neural network

phytoplankton

clustering 

Section on Statistics and the Environment