My research interest is in Artificial Intelligence (AI) with emphasis in developing single and multi-agent protocols to solve complex data-driven decision making problems. I am particularly interested in the interface between optimization (making better decisions), coordination (taking into account of agents' goals and preferences), and privacy (while protecting the individual's data from external attacks).

The word cloud on the right has been obtained on several recent abstracts of my papers.


Research Areas

Private AI

Every day, massive amounts of data are collected, shared, and used as input to machine learning and sophisticated optimization algorithms, providing valuable benefits to the society. However, sensitive properties can be inferred from these data, and inappropriate use might result in social and financial harm to the individuals. Managing privacy is a technical and societal challenge that must be carefully addressed to realize the promise of an ethical data-driven decision making.
Differential privacy (DP) is a property of an algorithm ensuring a strong privacy protection for individuals participating in a data analysis task. A DP algorithm protects an individual's data by injecting randomized noise into it. While such a process ensures privacy, it also impacts the quality of data analysis, creating a tradeoff between utility and privacy. Thus, a critical aspect of privacy research focuses on how to share sensitive data limiting disclosure while ensuring sufficient utility.
The main direction of my research in privacy is concerned with the protection of sensitive users' data for large-scale optimization and machine learning models while preserving the salient features of the application of interest.


Distributed AI

Multi-agent optimization is at the forefront of modern optimization theory and has undergone a dramatic development, stimulated by emerging applications, such as those powered by the Internet of Things, in a host of diverse disciplines. The Distributed Constrained Optimization Problem (DCOP) is a model in which multiple autonomous agents coordinate their decisions by exploiting a graphical model, and so to achieve a shared goal while accounting for personal preferences.
It can been used to model many multi-agent coordination problems, including scheduling of devices in a smart home, the coordination of autonomous vehicles, and the assignment of targets in sensor networks. The agents solve a combinatorial problem while the coordination process is executed within the boundaries of confined communication.
Although existing DCOP solvers offer desirable properties, such as using a bounded number of messages and localizing communication to between neighboring agents, their application to practical problems has been challenged due to their computational inefficiency and lack of generality. Thus, my research in multi-agent optimization has focused on increasing the efficiency and expressiveness of DCOPs.

Multiagent optimization

Scalable AI

A variety of central problems in AI, including finding the most probable explanation (MPE) in Bayesian networks and constraint optimization problems, are modeled as optimizing the costs of a network of cost functions. These graphical models are used to support decision making under uncertainty as well as to explain a domain, and are used in medical diagnosis, semantic search, and biological network learning. Despite their importance, the inherent complexity of these models challenges their scalability and usability in large complex problems.
Recently, massively parallel architectures, such as those found in Graphical Processing Units (GPUs), have witnessed an enormous success in a wide range of AI applications, including accelerated deep learning, image processing, and data analysis. A GPU offers thousands of computing cores and can provide an unprecedented application performance by offloading compute-intensive portions of the application into it.
Part of my research investigates how to devise efficient and scalable algorithms to solve large optimization problems over graphical models by exploiting the availability of GPUs.

Scalabe AI

Computational Biology

Computational Biology is a wide field of study which involves subjects as the study of Proteins Structure and of Gene Regulatory Networks. A Gene Regulatory Networks (GRN) is a collection of DNA segments in a cell which interact with each other either directly or indirectly (through their RNA and protein expression products). Understanding the behavior of GRNs is a core interest in computational biology and medicine, as it is closely related with the process of drug production and assessment. I have employed committee machines--that are techniques to aggregate multiple learners--to predict a GRN structure, and augment them with biologically meaningful constraints to filter irrelevant solutions.
Proteins are the product of coding genes, and perform a vast number of functions within living organisms. Their function is strictly related with their spatial conformation, therefore the Protein Structure Prediction (PSP) problem---which aims at predicting the 3D structure of a protein given its amino acid composition---is of central importance. We have developed a specialized Constraint Solver for the PSP problem, which employs efficient constraint propagators to search on the space of meaningful structures.

Computational Biology