Se ha denunciado esta presentación.
Se está descargando tu SlideShare. ×

AI at Google (30 min)

Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Anuncio
Cargando en…3
×

Eche un vistazo a continuación

1 de 37 Anuncio

AI at Google (30 min)

Descargar para leer sin conexión

Here's a high level overview of what motivates many AI teams at Google, what gives us confidence that humans will solve intelligence, the recent impact of advances in this work, and some examples of how people can get started today... for free! I first gave this talk to recipients of the 2019 AI for Good Awards, then again to recipients of the 2019 NASA FDL Challenge Fellowships. The slides are mainly a backdrop, but people still seemed to want a copy.

Here's a high level overview of what motivates many AI teams at Google, what gives us confidence that humans will solve intelligence, the recent impact of advances in this work, and some examples of how people can get started today... for free! I first gave this talk to recipients of the 2019 AI for Good Awards, then again to recipients of the 2019 NASA FDL Challenge Fellowships. The slides are mainly a backdrop, but people still seemed to want a copy.

Anuncio
Anuncio

Más Contenido Relacionado

Presentaciones para usted (20)

Similares a AI at Google (30 min) (20)

Anuncio

Más reciente (20)

AI at Google (30 min)

  1. 1. AI at Google Scott Penberthy Director of Applied AI, Google
  2. 2. Agenda Mission Examples How it works 1 2 3 4 Start Now!
  3. 3. Mission
  4. 4. Mission Solve Intelligence.1 2 Use (1) to solve everything else.
  5. 5. Approaches Hand-coded solutions Inspired by logic Learn from experience (data) Inspired by neuroscience
  6. 6. Framework
  7. 7. Inspiration https://physicsworld.com/a/quantum-microscope-peers-into-the-hydrogen-atom/
  8. 8. Confidence
  9. 9. Examples
  10. 10. Seeing Deep Learning has become a better driver than humans by literally going from pixels to steering, throttle, distance, and more. https://waymo.com/tech/
  11. 11. Hearing Deep Learning now powers real time audio speech translation for the top languages, allowing more humans to connect than before.
  12. 12. Speaking Deep Learning now generates realistic human speech, much how we evolved as mammals.
  13. 13. Reading Deep Learning can analyze 51 different file formats, condensing information to tensors for search, labeling and summarizing. Iron Mountain InSight™ 51 File Formats
  14. 14. Creating Generative techniques are now creating imagery and art with humans, as well as simulated environments for learning. source: thispersondoesnotexist.com
  15. 15. Operating Reinforcement learning techniques from AlphaGo are optimizing data centers, simulating quantum chemistry, and controlling robots. source: thispersondoesnotexist.com
  16. 16. How it works
  17. 17. Scalar Vector Matrix Tensor
  18. 18. input input 5 3 a b add mul c d add e 23 3 5 8 15 3 5 x F(x) Tensors flow through graphs
  19. 19. Gradient Descent - find F(x)
  20. 20. “Our results are 10^5-10^6 faster with double-digit process improvement...” ...multiple projects F(x) v. f(x) Universal Approximation
  21. 21. Tensor Pods (11.5 pflops)
  22. 22. Unity (US, JP) 2010 Monet (US, BR) 2017 Tannat (BR, UY, AR) 2017 Junior (Rio, Santos) 2017 FASTER (US, JP, TW) 2016 PLCN (HK, LA) 2019SJC (JP, HK, SG) 2013 Indigo (SG, ID, AU) 2019 Edge node locations >1000 Edge points of presence >100 Network Network sea cable investments The largest cloud network, comprising >100 points of presence 25% of the World’s Internet Traffic Tensor connectors - 1000x our speed
  23. 23. Impact? Product Users Data20161950 $ Cost of Prediction Source: “Managing the Machines: AI is making prediction cheap, posing new challenges for managers” by Ajay Agrawal, Joshua Gans, and Avi Goldfarb © 2016 (ajay@agrawal.ca)
  24. 24. Start now
  25. 25. Democratizing AI https://ai.google Normal humans AI Nerds ML frameworks: TensorFlow, XGBoost, Sklearn, PyTorch Cloud ML Engine: managed service for training & serving custom models RPA: Build robots for the office worker Kubeflow: deploy ML pipelines for pre-processing data, training and serving models on Kubernetes Deep Learning VM images: spin up VMs with popular ML frameworks pre-installed AutoML, BQML: train & serve no model code ML APIs: integrate AI into codebase
  26. 26. Process screens across SAP, Windows, Web, Citrix at 10+ clicks per second AI-powered Robotic Process Automation
  27. 27. Enable your entire team to automatically build and deploy state-of-the-art ML models on structured data at massively increased speed and scale. Cloud AutoML Tables
  28. 28. It’s almost not fair For each product: ● Relevant tables joined by given IDs ● Some minimal preprocessing done to match input requirements ● Run until converge ● Benchmarks run between H2 2018 to today (as they became available)
  29. 29. Learn through play https://ai.google/education
  30. 30. AIY Vision Build your own smart camera (cloud + edge)
  31. 31. AIY Voice Build your own assistant (cloud + edge)
  32. 32. ML5.js Tensorflow for poets https://www.youtube.com/watch?v=jmznx0Q1fP0
  33. 33. Text recognition Image labeling Barcode scanning Face detection Landmark recognition ML Kit Mobile SDK for vision, reading, ocr and barcodes
  34. 34. http://colab.research.google.com https://kubeflow.org AI Raspberry Kubes! Pico 3S Raspberry Pi Cluster, Coral Edge TPU
  35. 35. http://colab.research.google.com https://kubeflow.org (free) Personal Supercomputers
  36. 36. Exponential growth of AI Because tensors … work. Arxiv Papers 18 months Google Directories 18 months Model Computation 3.5 months
  37. 37. Thank you!

Notas del editor

  • And these boats are hard at work across the Pacific and the Atlantic oceans where we have been laying down the world's largest IP network, we have connections between every continent. (except antarctica)

    Blue is operational, Green is under construction, and you can see there is a region of world we are heavily focused on.
    Our network does not just connect our data centers to each other, like some others cloud providers, but connects us to nearly every ISP on the planet, and from a security and performance perspective this is amazing. Because we will be keeping your data on our network for high security and performance longer and closer to your customers.
    We control it entirely, end to end,
  • Limitations of Colab:

    You may hit the memory limit during training (~12 GB), will cause the runtime to start over
    May not be scalable for training jobs that take a long time
    Probably want a place to deploy your model in production after it’s been trained

  • [SARA]
    Switch to demo. Walkthrough doc is here: https://docs.google.com/document/d/1TuBlbheGuRrXqdiFvon8biFb_F9dbNO6KviX1fPAllA/edit
  • And based on benchmarks we’ve done, the results speak for themselves
    There are a number of vendors in this space, and we chose to benchmark against a subset of them with similar functionality
    Benchmarked on Kaggle competitions, which I love as a benchmark because they involve real data from a real company that is putting 10s to 100s of thousands of dollars of prize money on the line to get a good solution, and willing to wait months to get a result, and thousands of serious data scientists around the world compete
    X-axis, Y-axis
    Tables usually in the top 25% which is usually better than the existing vendors we tested. So overall, we do quite well
  • Limitations of Colab:

    You may hit the memory limit during training (~12 GB), will cause the runtime to start over
    May not be scalable for training jobs that take a long time
    Probably want a place to deploy your model in production after it’s been trained

  • [SARA]
    Switch to demo. Walkthrough doc is here: https://docs.google.com/document/d/1TuBlbheGuRrXqdiFvon8biFb_F9dbNO6KviX1fPAllA/edit

  • [SARA]
    Switch to demo. Walkthrough doc is here: https://docs.google.com/document/d/1TuBlbheGuRrXqdiFvon8biFb_F9dbNO6KviX1fPAllA/edit
  • And based on benchmarks we’ve done, the results speak for themselves
    There are a number of vendors in this space, and we chose to benchmark against a subset of them with similar functionality
    Benchmarked on Kaggle competitions, which I love as a benchmark because they involve real data from a real company that is putting 10s to 100s of thousands of dollars of prize money on the line to get a good solution, and willing to wait months to get a result, and thousands of serious data scientists around the world compete
    X-axis, Y-axis
    Tables usually in the top 25% which is usually better than the existing vendors we tested. So overall, we do quite well
  • Closing

×