Placeholder Image

Subtitles section Play video

  • KARTHIK KASHINATH: Extreme weather is changing.

  • There's more extreme rainfall, heavy flooding, forest fires.

  • There's the radio signature, [INAUDIBLE]..

  • Being able to predict these extreme events more accurately

  • is kind of the big challenge that we're facing right now.

  • There's 100 terabytes of climate data every day

  • from satellites, from observations, from models.

  • So climate data is a big data problem.

  • We need things that are fast that

  • can sift through all of that data rapidly and accurately.

  • And deep learning is almost perfectly poised for problems

  • in climate science.

  • THORSTEN KURTH: A lot of NERSC users are using TensorFlow.

  • It's one of the more popular frameworks.

  • We use TensorFlow to iterate quickly

  • over the different models, different layer parameters.

  • For this particular climate project,

  • to create the deep learning model,

  • we started from segmentation models,

  • which have proven to be successful,

  • for example, our satellite imagery segmentation tasks.

  • And then we use TensorFlow to enhance the models

  • until we found a set of models to perform well

  • enough for this specific task.

  • But for the volume of the data, complexity of the data,

  • the network required 14 teraflops.

  • So if you want to do this on your workstation,

  • it would take months to train.

  • MIKE HOUSTON: To really tackle these problems requires

  • the largest computational resources that

  • are available on the planet.

  • So systems like the Summit supercomputer,

  • it's two tennis courts in total size.

  • I mean, this thing is state-of-the-art.

  • It's a million times faster than your common laptop.

  • 3.3 exaflops.

  • Just imagine what you do at your workstation,

  • but now imagine having 27,000 times that power.

  • We can do that now.

  • THORSTEN KURTH: We were surprised how good it actually

  • scales.

  • 1,000 nodes, then 2,000 nodes.

  • 5,000 nodes.

  • MIKE HOUSTON: This was the first time

  • anybody's ever run an AI application at this scale.

  • Instead of having the climate scientists figure out

  • how to write high tune code, they

  • could express things in a very natural way in Python,

  • in TensorFlow, and get all the high performance code

  • that most HPC people are used to within TensorFlow.

  • KARTHIK KASHINATH: We're now entering the space where

  • AI can actually contribute to the predictions

  • of these extreme weather events.

  • MIKE HOUSTON: When you combine traditional HPC with AI,

  • you can tackle things we never thought that we could tackle.

  • Fusion reactor research, understanding diseases

  • like Alzheimer's, cancer, right?

  • That's incredible.

  • THORSTEN KURTH: We've shown that with the hyperactivity

  • framework such as TensorFlow, you can get to massive scale,

  • and you can get awesome performance

  • and accomplish your goals.

  • KARTHIK KASHINATH: Genetics, neuroscience, cosmology,

  • high energy physics, that is immensely exciting for me.

KARTHIK KASHINATH: Extreme weather is changing.

Subtitles and vocabulary

Operation of videos Adjust the video here to display the subtitles

B1 climate extreme data houston deep learning ai

Powered by TensorFlow: utilizing deep learning to better predict extreme weather

  • 3 0
    林宜悉 posted on 2020/03/25
Video vocabulary