Overview
This newer line of work uses simple models [bai25] to study aspects of machine learning (ML) that can be difficult to isolate in large architectures. The emphasis is on mechanisms rather than scale.
Questions include how stochastic dynamics, optimization rules, noise and representation spaces shape learning. Minimal models make it possible to compare intuition from statistical physics with behaviour observed in learning algorithms.
This theme is meant to remain flexible: it can include generative models, latent-variable approaches, learning dynamics and applications where simplified models expose the relevant structure.
Restricted Boltzmann machines are examples of simple, interpretable ML models. Their application to amino acid sequences in protein helices and sheets reveals what was termed the effective hydrophobicity of amino acids, which determines how Nature uses them in the hydrophobic patterns of proteins [bra23].
In addition, thanks also to the growing expertise of students in the master's program in Physics of Data at the University of Padova, the application of more complex ML models led to practical advances, for example in medical sciences [bra22,zan25] and the preservation of artistic paintings [cal25].
Selected papers
[bai25] Emergence of Nonequilibrium Latent Cycles in Unsupervised Generative Modeling
A simple generative model showing better performance when operating far from equilibrium, where it visits latent representations in cycles.
[bra23] Interpretable machine learning of amino acid patterns in proteins: a statistical ensemble approach
Restricted Boltzmann machines applied to study sequences of amino acids in proteins.
[bra22] Radiomics and deep learning methods for the prediction of 2-year overall survival in LUNG1 dataset
Study of a database of tumor images with various ML tools.
[zan25] xEEGNet: towards explainable AI in EEG dementia classification
This work introduces a compact and explainable neural network for electroencephalography (EEG) data analysis, which reduces overfitting through a major parameter reduction.
[cal25] Machine learning and numerical simulations for predicting critical crack conditions in wooden panels
A study of cracks in wooden panels that also uses tools from machine learning.