Public defence in Computer Science, MSc Prakhar Verma

Title of the thesis: Scalable Probabilistic Inference for Sequential Stochastic Models

Thesis defender: Prakhar Verma 
Opponent: Associate Professor Jes Frellsen, Technical University of Denmark
Custos: Associate Professor Arno Solin, Aalto University School of Science

Many real-world systems evolve over time under uncertainty. Examples include weather, financial markets, autonomous technologies, and scientific processes, where important aspects of the system cannot be observed directly and measurements are often noisy. To understand these systems and predict their future behaviour, machine learning (ML) models must reason about uncertainty rather than relying on a single best guess. 

This doctoral thesis develops new methods that enable ML models to learn from scarce and noisy data more efficiently and reliably. The work focuses on probabilistic models, which represent uncertainty mathematically and provide a principled way to estimate unknown quantities and predict future outcomes. Although these models are powerful, they often become too computationally demanding for complex real-world problems. 

To address these limitations, the thesis introduces new scalable methods that make probabilistic inference practical for increasingly complex sequential models. The proposed methods improve the analysis of continuously evolving systems, enable machine learning models to learn from data that arrive over time without storing all previous observations, and make it possible to train more expressive generative models efficiently. The research also develops a Bayesian approach for causal discovery that combines observational data with prior knowledge obtained from large language models, allowing cause-and-effect relationships to be identified more robustly even when both data and prior information are uncertain. 

Together, these methods show that probabilistic inference can be made both scalable and reliable, even under scarce and noisy data. The resulting tools are general: they let models learn continuously, quantify the uncertainty in their predictions, and apply across domains from scientific modelling to autonomous systems.

Keywords: Probabilistic Inference, Sequential Learning, Bayesian Machine Learning

Thesis available for public display 7 days prior to the defence at Aalto University's public display page. 

Contact Information: 
prakhar.verma@aalto.fi