Bachelor Student Prize Winner

Eighteen students completed their bachelor's degree in 2021 with a thesis that was labelled ‘excellent’ by their faculty. Here you will find a short introduction to these excellent theses in the form of an "elevator pitch" from each student, plus a video in which the supervisor briefly addresses the lucky winner.

Krzysztof Cybulski

 Faculty of Science and Engineering | Bachelor Data Science and Artificial Intelligence

"Reliable Deep Regression Using Overhead Images"

Krzysztof's elevator pitch
Deep neural networks are a family of powerful mathematical models which can predict high level information from raw data, such as images. They're particularly useful for researchers who wish to analyse a lot of images in an automated manner. Unfortunately, they're extremely complex and thus difficult to interpret statistically. There is a promising methodology - conformal prediction - which may alleviate this, as it allows for constructing predictive intervals on top of arbitrary models. In this thesis, I investigated the statistical properties of various configurations of deep neural networks with conformal prediction for the task of predicting neighbourhood statistics from aerial photographs. I discovered that conformal predictive intervals suffer from an unfortunate spatial bias, and looked into possible ways of correcting it.