Alessandro Manenti

We show that in deep learning models using latent categorical random variables (e.g., VAE, RL, GNNs), the softmax function - commonly used to parameterize the distribution - can hinder training. We theoretically motivate why this is the case using results from Information Geometry. We then propose an alternative parameterization, the catnat function, that - as we prove - has better information-theoretic properties. Extensive empirical evidence demonstrates its effectiveness across diverse settings.

We propose a linear-time approximation of Graph Random Features for computing graph kernels. We achieve this complexity by avoiding explicit graph materialization through the use of Gumbel-softmax sampling and random features. The approach is accelerator-friendly, theoretically grounded, and empirically shown to achieve order-of-magnitude speedups while matching or improving the accuracy of standard GRFs across synthetic benchmarks, on point clouds and on vision transformers.

Graph Neural Networks and related models benefit from relational information, but task-relevant relations are often latent. Accurately learning these latent structures and their uncertainty is difficult without direct supervision. In this paper we show that standard loss functions (e.g., MAE, MSE, ...), even when when in a probabilistic form, fail to recover such probabilistic relations. We prove and empirically show that a broad class of losses provides stronger guarantees into learning the correct uncertainty over the latent variables while maintaining predictive accuracy.

Ph.D. student

Phisics of Complex Systems - Master's Double Degree

Phisics - Bachelor's Degree

Lecture for USI — Master Program in AI (Graph Deep Learning course).

Lecture for USI — Master Program in AI (Advanced Topics in Machine Learning course).

Lecture for USI — Bachelor Program in Informatics (Machine Learning course).