Host
Florian Pappenberger
Speaker
Nal Kalchbrenner is a deep learning scientist and co-founder of the Google Brain team in Amsterdam. Nal has worked in numerous areas of deep learning with applications to a broad set of domains such as natural language understanding and translation (e.g. ByteNet), image and video models (e.g. PixelRNN), speech and audio models (e.g. WaveNet) and reinforcement learning for games (e.g. AlphaGo). Nal was previously a research scientist at Google DeepMind in London, after finishing a PhD in Computer Science at Oxford University, a MSc at the University of Amsterdam, and a BA/BS at Stanford University.
Abstract
In this talk we present MetNet, a neural network that forecasts precipitation up to 8 hours into the future at the high spatial resolution of 1 km2 and at the temporal resolution of 2 minutes with a latency in the order of seconds. MetNet takes as input radar and satellite data and forecast lead time and produces a probabilistic precipitation map. The architecture uses axial self-attention to aggregate the global context from a large input patch corresponding to a million square kilometers. We evaluate the performance of MetNet at various precipitation thresholds and find that MetNet outperforms Numerical Weather Prediction at forecasts of up to 7 to 8 hours on the scale of the continental United States. We will also discuss various properties of neural weather models in comparison to those of numerical weather prediction.