Postdoctoral Fellow
Research group | Statistics and Data Science
Main supervisor | Geir Olve Storvik
Co-supervisor | -
Affiliation | Department of Mathematics, UiO
Contact | ltronneb@math.uio.no
Short bio
I am a DSTrain Postdoctoral Fellow in machine learning and statistics in the department of mathematics. Prior to this I was a Research Associate (postdoc) at the MRC Biostatistics Unit at the University of Cambridge, and completed my PhD in Biostatistics at the University of Oslo in 2022. There I worked on probabilistic machine learning in the context of personalised medicine, modelling data coming from high-throughput drug screening in cancer, finding biomarkers of drug response and predicting treatment effects. My interests are broadly within the realms of probabilistic machine learning, and various connections to Bayesian nonparametrics, leveraging these connections to build novel models and shed light on old ones.
Research interests and hobbies
I’m interested in developing novel statistical methods that are accurate, can provide principled uncertainty quantification and can leverage domain expertise effectively. This is key to widespread adoption of machine learning models into e.g. biomedical science, where the datasets are small and noisy, but we know a lot about the underlying biological systems. In my DSTrain project, I’ll exploit connections between Bayesian nonparametrics and machine learning to study existing ML models, as well as develop new ones.
Outside of work, I enjoy cooking, good books and terrible movies.
DSTrain project
Bayesian adaptive networks via Kernels and Nonparametrics
Machine learning (ML) algorithms have revolutionised many scientific fields over the past decades. ML models enable powerful predictive modelling in increasingly complex datasets, and on data modalities that previously were considered out of reach of mathematical modelling such as images. However, traditional ML models often struggle in settings where the data is scarce or of low quality, do not provide uncertainty quantification of their predictions, and can tend to overfit the training dataset.
Bayesian machine learning has been proposed as a promising avenue to address these shortcomings, enabling practitioners to compensate for noisy or otherwise low-quality data by encoding relevant domain expertise in the prior, while the posterior provides consistent uncertainty quantification. Nonetheless, Bayesian Neural networks (BNNs) faces their own difficulties in practice, including difficulties in specifying sensible priors, computational complexity and architectural choices that impact the models’ performance.
BANKiN endeavours to advance the field of Bayesian Machine learning by leveraging the deep connections between Gaussian Processes (GPs) and BNNs, using the more interpretable GPs to inform prior decisions and construct surrogate models that combine the best features of each approach. The project also makes use of modern Bayesian nonparametric (BNP) methods to inform architecture design in deep network models, allowing these models to automatically adapt to the complexity of the modelling task and avoid overfitting the training data.
Through theoretical analysis, empirical benchmarking, and the development of novel Bayesian models, BANKiN seeks to enhance model interpretability, improve scalability and efficiency, and enable flexible model complexity in deep network models, extending the reach of these models into application domains such healthcare and biology.