The recent wave of machine-learning based Artificial-Intelligence technologies is having a huge societal and economic impact, with AI being (often silently) embedded in most of our everyday experiences, such as virtual assistants, tracking devices, social media, recommender systems.

Every year, each of us leaves behind several gigabytes of “digital breadcrumbs”, overspread in disparate systems that we use in our daily activities. Making sense of this data requires appropriate retrieval, aggregation and analysis. Currently the dominant paradigm of machine-learning -based AI is centralized, as users interact with service providers and the data generated at each interaction are stored into large, secluded databases subsequently used to train machine-learning models.

This has multiple drawbacks, including:

• state-of-the art uses of massive big data collections enable AI systems based on Deep Neural Networks which are often black boxes, non-interpretable even for experts, making them opaque and hardly trusted;
• the aggregated information collected is mostly used to benefit private interests and not the public (e.g. for raising collective awareness and improving social wellbeing);
• each individual’s data make up a virtual jigsaw puzzle whose pieces are scattered across different, closed-off, centralised platforms, making it hard for individuals to reconstruct a comprehensive and yet private picture of themselves;
• collecting all data in central locations might become technically unfeasible, due to the sheer amount of data generated “at the edge” of the Internet, and due to constraints in data mobility resulting from data ownership requirements.

SAI, a decentralized, collective and explainable AI

The research community (and society in general) has already realised that the current centralised approach to AI is not an acceptable and sustainable model in the long run.

The “next wave” of machine-learning driven AI should be human-centric, explainable, more distributed and decentralised.

These principles address the societal and ethical expectations for trustworthy, privacy-respectful AI, such as those recommended at the European Level ( Ethics Guidelines for Trustworthy AI). They also fit a clear trend to develop decentralised machine-learning for strictly technical reasons: performance, scalability, real-time constraints.

The vision of SAI is towards a decentralised “collective” of local machine-learning -based AI components interpreting data and interacting according to human-centric design principles, where explainability is guaranteed both at the local and collective level.

SAI will develop the scientific foundations for novel machine-learning -based AI systems ensuring

• individuation: in SAI each individual is associated with their own “Personal AI Valet” (PAIV), which acts as the individual’s proxy in a complex ecosystem of interacting PAIVs;
• personalisation: PAIVs process individuals’ data via explainable AI models tailored to the specific characteristics of their human twins;
• purposeful interaction: PAIVs interact with each other, to build global AI models and/or come up with collective decisions starting from the local models;
• human-centricity: novel AI algorithms and the interaction between PAIVs are driven by quantifiable models of the individual and social behaviour of their human users;
• explainability: explainable machine-learning techniques are extended through quantifiable human behavioural models and network science analysis to make both local and global AI models explainable-by-design.