Malicious use of AI
My main focus at the moment is techniques for mitigating or preventing malicious use of neural network-based AI systems. This includes malicious applications and hacking of AI systems (using e.g. data poisoning or adversarial examples).
Keywords: AI safety, machine learning, neural networks.
Improving classification methods and predicting one-sided-violence events
Using one of the world’s best databases on conflict data, the aim of this project is to make forecasts for one-sided-violence events, defined as the use of armed force against civilians, either by the government or by an organised group. This involves developing classification methods that deal with methodological challenges such as class imbalances in the data and how best to assign weights to different variables. The project is a collaboration with PhD student David Randahl at the Department of Peace and Conflict Research at Uppsala University, whom I co-supervise.
Keywords: classification, prediction, machine learning.
Analysis and modeling of data from automated milking systems
Automated milking systems (AMS) are an integral part of modern high-tech dairy farms. For each milking, AMS collect thousands of data points through sensors and recording systems. The aim of this project is to develop methods for analysing the massive high-frequency datasets that result, covering hundreds of thousands of milkings. This includes automated analysis of milk flow during milking and total milk yield over time, with the end-goal of improving animal health while maintaining a high milk production. The project is a collaboration with industry and researchers at the Beijer Laboratory for Animal Science and the at the Swedish University of Agricultural Sciences.
Keywords: big data, complex data, functional data analysis, machine learning.
Methods for handling censored data
Measurements are said to be censored if it only is known that they are above or below some limit, but not by how much. Traditionally, censored data arises in studies involving the time until an event, where the event may not have occured for all individuals when the study ends. In such cases, we only know that the time until the event is greater than the length of the study. Censoring has also become a frequent issue in modern medicine and biology, where concentrations of chemicals are measured e.g. in samples of body fluids. When a measurement falls below the detection limit of the instrument used, the exact concentration of the chemical remains unknown. This project is concerned with evaluation and development of methods that handle censored multivariate data. It is motived by applications in medicine and proteomics, most of which come from my continuing collaboration with researchers at the Uppsala Berzelii Technology Centre for Neurodiagnostics.
Keywords: censoring, hypothesis testing, estimation, classification.
Other research interests
In addition the topics listed above, my research interests include:
- Multivariate statistics, in particular hypothesis testing, classification and high-dimensional problems,
- Inference for discrete distributions, such as inference about proportions or odds ratios,
- Computational statistics,
- Invariance properties of statistical methods,
- Asymptotic expansions,
- The interplay between Bayesian and frequentist ideas.
See also my list of publications.