Publications
Peer-reviewed papers, preprints, and theses. See also Google Scholar.
Peer-reviewed
-
Building on insights from fuzzy logic and binarized neural networks, we developed methods that significantly improve the performance of Physics-inspired graph neural networks (PI-GNNs) on dense combinatorial optimization problems. Our research found that PI-GNNs struggle with denser problems due to their dependency on phase transition during training. We resolved the issue by informing the network about the relaxation of the underlying binary problem, adjusting the activation function and the loss.
Paper Preprint (arXiv) -
A novel deep learning approach leverages the Rubik's Cube's inherent symmetries to solve it efficiently. By thoroughly analyzing the problem's group-theoretical properties, we developed a symmetry-invariant neural architecture with graph convolutional layers that significantly compresses the cube's vast state space. This new architecture demonstrates superior performance, enhanced generalization, and improved problem-solving abilities compared to previous state-of-the-art results.
Paper -
We propose a governance-aligned framework for evaluating explainable-AI methods, drawing on key desiderata from recent XAI research. The framework assesses explanation techniques along dimensions such as stakeholder understandability, faithfulness, verifiability, actionability, and scalability. By linking these criteria to practical oversight and audit needs, including those emerging under the EU AI Act, the approach helps institutions select XAI methods that integrate effectively into organizational decision processes and compliance workflows.
Preprint (PDF) Event report
Preprints
-
A philosophical critique of causal machine learning that applies Ordinary Language methodology to investigate the semantics of cause and effect reasoning. The work examines how causality manifests differently across scientific domains (physics/engineering, biology, social sciences) and argues that definitive causal claims require evidence aggregation across multiple domains, highlighting methodological challenges in harmonizing language games and emergence across scales.
Preprint (arXiv)
Theses
-
A contribution to the Geometric Deep Learning framework by studying symmetries in planning problems and implementing symmetry-invariant graph neural network for the Rubik's cube problem. The presented architecture outperforms previous state-of-the-art solutions in cost-to-goal regression experiments and exhibits high interpretability, with explanations based on intrinsic properties being preferable to standard explainer algorithms.
Thesis -
A comprehensive exploration of neural architectures that explicitly exploit domain symmetries, including weight-sharing schemes, convolutions, Deep Sets, and graph neural networks. The work demonstrates that symmetry-aware architectures outperform conventional models on problems exhibiting various symmetries, from simple functions like XOR to complex tasks like combinatorial problems.
Thesis