Machine learning for robust and reliable measurements

In this project, we will employ deep learning techniques to enhance the accuracy, interpretability, and robustness of indirect measurements. Particularly, we focus on problems of i) interpretation and analysis of metagenomic data obtained from agricultural soil samples, characterized by high-dimensional feature spaces with a relatively small number of soil samples (for an overview, see [1]-[4]), and ii) improvement of the non-invasive measurement approach developed for estimating animal weight based on 3D images in the farming industry, involving a huge amount of data used for volumetric representations (for a brief review, see [5]-[8]).

Both problems involve evaluating or detecting an unknown quantity from the observations indirectly related to the measured/detected quantity, and our goal is to apply low-rank models (which are increasingly relevant in data science) to both problems, given their ability to capture the essence of complex data while reducing the dimensionality. So, we are mainly interested in two major benefits they offer. One is related to overfitting reduction (the limited dimensionality confers a greater capacity for generalization, making them less prone to overfitting), and the second is related to the adaptation capacity, which is particularly important in real-time data processing.

Faculty Supervisor:

Leszek Szczecinski

Student:

Partner:

Universidade Tecnológica Federal do Paraná

Discipline:

Computer science

Sector:

Education

University:

Université du Québec : Institut national de la recherche scientifique

Program: