Samuel Cognolato

I received my bachelor's degree in Information Engineering in September 2019, and my master's degree in Computer Engineering in December 2021, both at the University of Padua, Italy. In my master's thesis, I explored how to build and train a neural network model that could perform elementary addition with an arbitrary number of operands and digits. This convolutional transformer-based model can learn an algorithm for addition end-to-end, just by observing the training examples, without intermediate steps, revealing an interesting process and internal representation of digits.

I was always intrigued by how to bestow reasoning and learning capabilities upon a model, having as little prior knowledge as possible. To do this, I'm looking into graphs as a way to represent knowledge, and Graph Generative Models as a means to build them from several sources of information.

High-level research

My high-level research topic regards integrating multimodal information through graphs. Sensory signals are what enables humans to experience the outside world. The encoding of each sensory channel is referred to as a sensory modality, and the human brain is capable of combining them. How to endow machines with analogous cognitive capabilities is still an open problem, studied in the discipline of multimodal machine learning, and leads to a wide variety of applications: from audiovisual speech recognition to text-to-image generation, the latter recently attracting widespread public attention. Still, the interactions between pieces of information from different modalities are not explicitly modeled or clearly interpretable. My aim is to build a bridge between heterogeneous sources of information with graphs, a mathematical tool explicitly representing objects and their relationships. I will focus on the generation and manipulation of these structures, serving as common ground for gathering and producing content from/to different modalities. To do so I will explore the field of deep generative models for graphs, with the aim of conditioning such processes using multimodal information.

 

Current focus

My current focus is on Deep Graph Generative Models (DGGM), that is, deep learning models capable of generating new graphs by mimicking those from a given dataset. More precisely, the model learns the probability distribution of the graphs in a dataset, from which we can sample.

• Deep Graph Generative Models


• Multimodal integration

1. S. Cognolato and A. Testolin, Transformers discover an elementary calculation system exploiting local attention and grid-like problem representation, International Joint Conference on Neural Networks (IJCNN), Padua, Italy, 2022, pp. 1-8, doi: 10.1109/IJCNN55064.2022.9892619.


2. S. Cognolato, A. Sperduti, and L. Serafini, IFH: a diffusion framework for flexible design of graph generative models, European Conference on Artificial Intelligence (ECAI), 2024, pp. 3039-3046. IOS Press. doi: 10.3233/FAIA240845


3. S. Cognolato, D. Rigoni, M. Ballarini, L. Serafini, S. Moro, A. Sperduti. D4: Distance Diffusion for a Truly Equivariant Molecular Design, 33th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), 2025. In press.