Mattia Jacopo Villani

CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching (2026): use MMD herding algorithm to improve performance of prior fitted networks. I helped with the implementation and experiments and co-devised the algorithm with Jamie.

Entropy Distribution as a Fingerprint for Hallucinations in Generative Models (2026): a new method to detect hallucinations in generative models. I found the research direction and got amazing help to get this to work by my colleagues.

MetaTT: A global tensor-train Adapter for parameter-efficient fine-tuning (2025): you might have used LoRA before. Here we propose a new method for parameter-efficient fine-tuning. High parameter compression compared to LoRA (~10x) and comparable performance using tensor train decomposition. I helped with some engineering here.

A Unified Framework for Provably Efficient Algorithms to Estimate Shapley Values (2025 - NeurIPS): there are many ways to estimate Shapley values (an important feature importance technique). Two of the main approaches are matrix-vector and least-squares estimation. We look at unifying these approaches and provide a general framework for understanding Shapley estimation. I developed the algorithmic implementation, which now works for high dimensional problems, and performed experiments.

Any Deep ReLU Network is Shallow (2025 - ECAI): Nandi and I showed that you can write a network with $L$ layers as a network with 3 layers, if you allow your weights to be in the extended reals. The proof is constructive and has some interesting implications for interpretability and pruning.

Relating Piecewise Linear Kolmogorov Arnold Networks to ReLU Networks (2025 - AISTATS): Nandi, Niels and I had a look at how to relate KANs and ReLU Networks. Turns out we can write any ReLU network as a KAN and vice versa. We all developed the theory together.

PICE: Polyhedral Complex Informed Counterfactual Explanations (2025 - AIES in AAAI): we can get exactly minimal counterfactuals for ReLU networks, with sufficient compute. This was my internship project.

Trading-off accuracy and communication cost in federated learning (2025 - AAMAS): using training by sampling, we can get incredibly cheap communication costs in federated learning. There’s also a nice connection to Zonotope geometry.

Graph Convolutional Neural Networks as Parametric CoKleisli morphisms (2022 - SYCO-10 workshop): Bruno and I sat down to look at GCNs via category theory. This paper was the result of a weekend together in London. The paper just describes weight sharing in GCNs. Bruno presented it nicely here.

Feature Importance for Time Series Data: Improving KernelSHAP(2022 - ICAIF workshop): using Shapley value as a feature importance techniques relies on the assumption that every feature is a player in a game. If our model is a time series model, this isn’t quite right. We propose a method for computing Shapley for time series. This was my internship project.

United Italian Societies (2017-2024): I helped Umberto set up this non-profit organisation, back in 2022. Before that, I managed a precursor project called Italian Societies Student Network. Within the UIS I founded the Research Centre. We helped Italian students in the UK connect, orgnaise educational events and provided support.

The Red Flower Factory (2020): I helped set up this spin-off start-up from Sevendots - a CPG marketing consultancy.