OSMOSIS: Open-Source Multi-Organizational Collaborative Training for Societal-Scale AI Systems

The goal of our project is to develop a novel framework and cloud-based implementation for facilitating collaboration among highly heterogeneous research, development, and educational settings. Currently, AI models for real-world intelligent systems are rarely trained as part of a collaborative process across multiple entities. However, collaboration among different companies and institutions can increase AI model robustness and resource efficiency. Towards a more efficient development process of AI systems at massive scale, we propose a general framework for AI model sharing and incentivization structures for seamless collaboration across diverse models, devices, use cases, and underlying data distributions. Through distributed sharing of AI models in a secured, privacy-preserving, and incentivized manner, our proposed framework enables significant cost reduction of system development as well as increased system robustness and scalability.

Caption for picture: Our proposed Open-Source Multi-Organizational Collaborative Training for Societal-Scale AI Systems design. Each agency/company uses its data to generate and update its own model (A). OSMOSIS, installed in each company data center, uploads the latest model update (B) to the common benchmark and collaborative model update services (C). When an update improves the performance of the collaborative model, OSMOSIS distributes the collaborative model update to the constituent companies (D) to update their models to the current state-of-the-art.

Check out more about this project on the Red Hat Research website.