In 2024, we have ample evidence that foundation models constitute a paradigm shift in artificial intelligence, demonstrating remarkable capabilities across a wide range of tasks. However, the true potential of FM lies in their ability to generalize across diverse domains and modalities, a frontier that remains largely unexplored. DVPS pushes this frontier by advancing the science and technology of multimodal foundation models (MMFM). Diversis Viis, Plurima Solvens: in English, "various ways, solving many", meaning "Various modalities, to solve many tasks".

DVPS envisions a future where the design and implementation of MMFM grows from an art based on heuristics and domain-specific tricks to a rigorous science. We aim to:

● Develop generalizable methods that work across diverse modalities and application domains.

● Create a unified framework for MMFM development, incorporating formal characteristics of each modality

● Establish principled approaches for integrating new modalities into existing FM and expanding the multimodal capabilities of current models.

DVPS considers a diverse environment for development, encompassing three substantially different core application domains: Cardiology, Geo-intelligence, and Language Communication. Furthermore,  the concept of “surprise application domains” that will be added during the project, to force our methods to generalize beyond our initial assumptions, driving innovation in MMFM science.