Our long-term goal is to obtain machine-intelligent modeling systems that seamlessly integrate stochastic ocean dynamical models and their multi-fidelity representations with Bayesian and generative learning from data-model misfits, to construct improved ocean models with more accurate parameterizations and discover invariances or differential equations, over a range of spatial and temporal scales. We will use and vastly extend stochastic dynamic reduced-order methods, Bayesian GPs, adaptive DL schemes, and generative networks. Ultimately, our symbolic interpretation of ML models into emergent dynamical and constitutive relations would further compress knowledge relative to deep networks, thus extending outside the range of the training data.