The ultimate goal of SAI is to provide the foundational elements enabling a decentralised collective of explainable PAIVs to evolve local and global AI models, whose processes and decisions are transparent, explainable and tailored to the needs and constraints of individual users.
This goal is pursued through the following steps.
Develop a PAIV, integrating user’s data and explainable machine-learning models. The objective is to drastically extend current Personal Data Stores (PDS), mostly focused on access to data towards local proxies of their users, integrating explainable ML components and ways to share data/models with other PAIVs.
Design human-centric, collective machine-learning-based AI algorithms. The objective is to design algorithms for collaborative interaction between PAIVs’ machine-learning models, incorporating quantitative models of users’ individual and social behaviour.
Explain globally and in a measurable manner. Collective algorithms must preserve
the transparency and explainability of the local models. This means that:
• the global machine-learning models obtained, e.g., via federated learning, must remain explainable
• the collective decisions of PAIVs must be transparent to the users
• the quality of explanations must be measurable
Model the emergent behaviour at the collective level. We will study the emergent behaviour of collective
AI through network science modelling techniques and large-scale simulation. The objectives are:
• to derive compact analytical models capturing the emergent collective behaviour of PAIVs
(seen as a network of linked nodes)
• to develop scalable social simulation tools to study PAIVs interactions on a large scale
Validate the effectiveness of SAI in three specific scenarios (private traffic management, information diffusion in Online Social Networks (OSNs), and pandemic tracking and control). We will use available datasets and collect new ones to define the “as-is” benchmark with current machine-learning approaches, and to exploit analytical models and simulation tools to study the “to-be” scenario with SAI algorithms.