In this research, we develop a state-of-the-art GPU parallel implementations of two classical statistical models - the Lasso and multinomial logistic regression. To that end we also develop novel parallel implementations for matrix multiplication, Cholesky decomposition, and sampling from Polya-Gamma distribution. We achieve speedups of up to 100-fold compared to CPU versions.
COBISS.SI-ID: 1537467843
In this study we provide insight into how air pollution in Slovenia has changed during a 16-year period. In particular, we demonstrate how changes in climate, if not taken into account, can bias such analyses towards overestimating how much of the decrease in pollution is due to lower emission levels.
COBISS.SI-ID: 1537827267
This work addresses two problems in the area of parallel random number generation. First, we implement a library with over 20 parallel random number generators that can be easily used and is not specific to hardware from a particular vendor. And second, we provide a first broad study of which parallel number generators are the most efficient and most robust in terms of the statistical properties of the sequences of random numbers they generate.
COBISS.SI-ID: 1538103747
The problem of predicting air pollution levels is important for health reasons and required by EU legislation. It can be set as a point-prediction task, a task of predicting levels with some measure of uncertainty, or what is most common for legislative reasons, predicting whether predictions will rise over a certain threshold. With a Bayesian approach, we cast all these tasks as a single task of density prediction and all decision making is moved outside of the modelling process. The key practical advantage of this is that models do not have to be changed if the legislation, thresholds, or any other part of the decision making process changes. We demonstrate the methodology on several years of air pollution data from Slovenian measuring stations.
COBISS.SI-ID: 1537745603
We develop a novel Bayesian approach to modelling large-scale rubric-based assessment results where multiple raters rate participants but not each rater rates each participant. We apply the approach to Slovenian Matura Exam Slovenian language Essay. Results show that rater unreliability and bias have a practically significant effect on student scores. Most of the unreliability of scores is due to rater unreliability and not their bias.
COBISS.SI-ID: 1537763779