Functional Generative Networks
A way of generating varied but physically sensible predictions by adding randomness inside the network rather than to the output.
When you want a model to show you several possible futures, not just one, you have to inject randomness somewhere. The usual approach, used by diffusion models, sprinkles noise across the output itself and then cleans it up over many passes. It works, but it is slow, and the variations you get can look like noisy copies of the same answer.
Functional Generative Networks put the randomness into the network’s own control layers instead. The result is a single pass that produces genuinely different scenarios which each still obey the rules the model has learned, so a set of weather forecasts comes out as several plausible storms rather than one blurry one. Google DeepMind used the method in WeatherNext Cyclones, where it runs about eight times faster than the diffusion approach it replaced.