Theoretical & applied ML

Research

The SML lab carries out research on different topics of theoretical and applied Machine Learning, with a focus on domains characterized by complex structures like sequences, trees, graphs and relational knowledge bases. Our main research activities involve combining statistical and symbolic approaches to learning, integrating logic and constraint programming with statistical learning.

Neuro-Symbolic AI & Reasoning Shortcuts

Integrating logic and background knowledge with neural models, and studying when and why neuro-symbolic predictors learn the "wrong" concepts while still solving the downstream task (reasoning shortcuts) — including benchmarks and awareness techniques.

Structured-Output Prediction & Learning to Optimize

Predicting and optimizing over combinatorial structures — sequences, trees, graphs — with constrained and declarative learning approaches.

Reasoning & Learning in Hybrid Domains

Combining continuous and discrete/symbolic reasoning, weighted model integration, and probabilistic inference under constraints.

Explainable & Interactive Machine Learning

Interpreting large language models and deep networks, human-machine collaborative decision-making, and algorithmic recourse for overturning unfavourable automated decisions.

Constructive Preference Elicitation

Interactive methods for eliciting user preferences and guiding people toward effective, personalized recommendations and decisions.

Bioinformatics Applications

Applying structured and relational machine learning methods to problems in computational biology and microbiology.

Uncertainty Quantification & Learning to Defer

Calibration of machine learning classifiers and learning-to-defer frameworks that let a model hand off a decision to a human expert when appropriate.

Geometric Deep Learning

Graph Neural Networks across several fronts: explainability, self-explainable and domain-invariant architectures, temporal and link-level representation learning, and neuro-symbolic GNNs that combine relational reasoning with learned representations.

See the Publications page for the full, auto-updated list of papers across these areas, and Funding for the funded research programmes that support this work.