Atmosphere Ocean Science Friday Seminar

---

Capturing Long-range Fluid Interactions in "Shallow" Neural Nets

Speaker: Michael Lever, CAOS

Location: Online

Date: Friday, April 23, 2021, 4 p.m.

Notes:

In data-driven fluids modeling, a common goal is to train a neural network as an operator on the state of the system. Common targets include time tendencies and future states. Long-range interactions can be important in learning these targets. Traditional convolutional neural networks (CNNs) can capture long range interaction by using many layers and repeated pooling. However, a net like this is a dark grey box that needs 1E5-ish weights to capture the various flow regimes. I’ll talk about some multiscale and nonlinear (in the input) approaches to create smaller, more transparent, and human-readable neural networks.