Thesis subject

MSc thesis topic: Feature elimination for pasture growth models with Bayesian Neural Networks

While neural networks can give good prediction accuracy, they often behave as black boxes, which makes it hard to determine which parts of the input data are essential for making predictions. In many agricultural applications farmers can only afford a limited number of sensors.

It is therefore desirable to use a neural network architecture which can identify uninformative input data. This feature elimination is useful as it can indicate which measurements don’t need to be performed by farmers. A bayesian neural network with a tied horseshoe prior distribution for the weights of the first layer can automatically eliminate input data[1].

The purpose of this project is to investigate the use of a horseshoeBNN for predicting the nitrogen response in pasture growth. The dataset consists of simulations of pasture growth on eight sites in New Zealand [2]. The student will first train a Bayesian neural network with static input features, by reducing the time series data to summary statistics (mean, min, max etc.). The student will then develop an extended model architecture which uses the entire time series. The feature elimination results provided by the models will be compared to previous findings [3].

Objectives

  • Develop a deep learning pipeline with static input features
  • Adapt the model to handle time series data
  • Evaluate the deep learning approach in a case study

Literature

  1. Interpretable Outcome Prediction with Sparse Bayesian Neural Networks in Intensive Care, H. Overweg, A. Popkes, A. Ercole, Y. Li, J. M. Hernández-Lobato, Y.Zaykov, C. Zhang, HSYS (2020)
  2. Location-specific vs location-agnostic machine learning metamodels for predicting pasture nitrogen response rate C. Pylianidis, I. Athanasiadis, V. Snow, D. Holzworth and J. Bryant, ICPR 2020 Workshop "Machine Learning Advances Environmental Science (MAES)"
  3. Emulation of pasture growth response to nitrogen application, I. Athanasiadis, J. Bryant, D. Holzworth, C.Pylianidis, V. Snow, Internal document

Requirements

  • FTE-35306 Machine Learning
  • GRS-34806 Deep Learning

Theme(s): Modelling & visualisation