Contents:
Links:
- Part 1: Jeff Dean
- Part 2: Panel discussion 1 (Yoshua Bengio, Andrew Ng, Carla Gomes, Lester Mackey, Jeff Dean)
- Part 4: Panel discussion 2
Invited Talks
Jeff Dean: Computation + Systems vs Climate Change
ML is the challenge for the 21st century.
Getting to zero-carbon computation for Google:
- Focus for 6+ years; from 2012 to 2017 ramped up to 100% renewable energy purchase; now net-zero carbon* where the * is that it's only averaged over time.
Computation and ML to help fight climate change. Dean's general approach is Algorithm → System → Climate Application
- Bayesian inference → TensorFlow Probability → Fusion energy. Examples:
- Fusion energy models at TAE where they make vibrational model for Norman machine to estimate plasma conditions w/ TF Probability
- Also: Forecasting greenhouse gas concentrations, forecasting demand for energy or electricity, predicting the occurrence and spread of forest fires, detecting deforestation and environmental change, modelling salinity in estuaries and coral reefs
- Simulation → TPUs → Flood forecasting
- Google has hydraulic model maps that are 90x the resolution of public maps, but that'd be 700,000x as much compute to simulate; TPUs can do these computations efficiently enough to make them realistic.
- Google has alerted 2B people over 100k's of events (including floods) through Maps and Android.
- (Simulation → Neural Network Proxies) → TensorFlow → Weather Prediction
-
Current weather models use partial differential equations (PDEs) that require a large amount of computation especially as compute increases when you make your grid finer; EU weather just switched to 9km grid, while 1km grid is desired for cloud simulation.
-
Or use a neural network directly from sensor predictions, which is more efficient and makes a more fine-grained (in time and space) prediction at the short term (< 12 hours)
Technology + People—leverage: What information do people need to make good decisions? What are the information leverage points? Who should we build tools for?
- Environmental insights explorer: estimating building emissions, transport emissions, rooftop solar potential (e.g. Project Sunroof)
- Challenge to community: building climate decision-making tools using AI/ML for all the high leverage points.
Questions:
- Why not put servers in places where all electricity is renewable (e.g. Quebec with hydro)? → This is a big factor in where Google decides to build a new datacenter.