Sai Ravela presents at Shell, arguing for theory-data symbiosis, which includes:
- Models from theory suffer from nonlinearity, dimensionality and uncertainty
- Models from data suffer from the above plus issues of generalization, extrapolation, and the representation of theory
- Can we do better than either source alone? Three ideas:
- Machines complied from theory and adapted with data — using models to accelerate learnng
- Using theory to configure RKHS, stabilize and constrain learning
- Using learning to accelerate solutions to inverse problems
- Optimizing stability of hybrid systems
- New class of Neural Flow Kernels