Who we are

We are a dynamic research group at the University of Turin dedicated to exploring advancements at the intersection of computational biology and neural networks. Our team comprises passionate researchers driven by a common goal: to unravel the complexities of biological systems through innovative computational methods.

Computational Biology

In our computational biology division, we leverage advanced computational techniques to study gene regulation mechanisms. By integrating bioinformatics, statistical modeling, and machine learning, we aim to decipher the complex networks governing gene expression. Additionally, we investigate scale laws that govern biological systems, seeking to uncover universal principles underlying biological organization. Furthermore, our research delves into the intrinsic dimensions of biological datasets, employing dimensionality reduction and manifold learning methods to extract meaningful insights from complex biological data.

Neural Networks

Our neural networks section is dedicated to theoretical explorations of artificial intelligence. We believe that it is crucial to focus our efforts on the theoretical aspects of neural networks, given that their development relies not on a deep understanding but on trial and error. We believe that taking tools from physics such as statistqical mechanics and dynamical systems theory and applying it to neural networks can help us understand the underlying principles of these systems.

Topic Modeling

Topic modeling is a popular method for uncovering important patterns within vast datasets, connected the task of identifying hidden structures in data to network theory's community detection problem. This linkage has led us to developing innovative topic modeling techniques aimed at addressing shortcomings of traditional methods.

We showed how to recover information about the underlying structure of gene expression data by applying topic modelingin in our recent research analyzing TCGA breast and lung cancer transcriptomic data (Valle et al., 2020). This study successfully applied advanced topic modeling techniques to uncover latent structures within gene expression patterns associated with cancer subtypes. The findings showcase how topic modeling can reveal biologically relevant information, such as gene-enriched topics linked to disease characteristics and patient survival probabilities.

Stragglers

Understanding neural networks can be challenging due to how they learn. The process involves navigating a complex landscape to find optimal solutions in high-dimensional spaces. The goal is to avoid getting stuck in less-than-ideal outcomes while converging towards broadly applicable solutions, which helps overcome the curse of dimensionality.

To explore this, we examined neural network structures and uncovered interesting dynamics in how information is processed. We observed a pattern of compressing and then decompressing hidden representations during learning. Certain data points, named stragglers, influence the network's behavior in unexpected ways, leading to non-linear learning trajectories (Ciceri et al., 2023).