Democratize ai with google cloud
2 likes563 views
The document discusses how Google Cloud is democratizing AI to make it accessible to a wider range of users beyond data scientists, emphasizing the integration of AI across various applications and platforms. It highlights tools such as Cloud AutoML and TensorFlow for training and deploying machine learning models, as well as the use of edge TPU for real-time analytics. The key focuses are speeding up the AI lifecycle, embedding AI everywhere, and facilitating collaboration amongst developers and practitioners.
Democratize ai with google cloud
- 1. Machine Learning for everyone with Google Cloud Henrik Hammer Eliassen Customer Engineer Google Cloud henrikhammer@google.com
- 2. Democratize Your AI Speed up Your AI lifecycle Speed up and support all the steps in the full AI lifecycle and seamlessly integrate between them. Democratize AI by making it available to a broad range of skills backgrounds - not just data scientists Train and Serve your AI everywhere so it becomes more useful for the user 1 2 3 Three directions to make AI more useful Embed Your AI everywhere
- 4. 1000’s Deep learning researchers Who can actually use AI today? 21M Developers ~1M Data scientists Very few users today can create a custom ML model We need to make AI accessible to millions more
- 5. AI on Google Cloud Platform Application developers Machine Learning APIs Cloud AutoML Cloud ML Engine Data scientists & ML practitioners new
- 6. AI on Google Cloud Platform Application developers Machine Learning APIs Cloud AutoML Cloud ML Engine Data scientists & ML practitioners new
- 7. Use a pre-trained model to accomplish common AI tasks Cloud Vision Cloud Natural LanguageCloud speech to text & text to speech Cloud Video Intelligence Cloud Translation
- 8. Cloud Vision API Label & web detection OCR Logo detection Explicit content detectionCrop hintsLandmark detection
- 9. Cloud Vision Example Application developers Machine Learning APIs Cloud AutoML Cloud ML Engine Data scientists & ML practitioners new http://vision-explorer.reactive.ai/#/galaxy?_k=t4xe0d
- 10. Cloud Vision API Image Request Response
- 13. AI on Google Cloud Platform Application developers Machine Learning APIs Cloud AutoML Cloud ML Engine Data scientists & ML practitioners new
- 14. Dataset Cloud AutoML Train Deploy Serve Generate predictions with a REST API Cloud AutoML for Vision, NLP, Translation ML that creates ML for your problem
- 15. I want to predict weather trends and flight plans from images of clouds AutoML Vision Demo
- 16. Let’s try the Vision API
- 17. Let’s try Auto ML https://youtu.be/gVz9jKE_9iU?t=920
- 18. Wait a minute !!! How good is an AutoML model?
- 19. Inception-ResNet-v2 Years of effort by top ML researchers in the world computational cost Accuracy(precision@1) accuracy Learning Transferable Architectures for Scalable Image Recognition, Zoph et al. 2017, https://arxiv.org/abs/1707.07012 Handcrafted models by cost and accuracy
- 20. Learning Transferable Architectures for Scalable Image Recognition, Zoph et al. 2017, https://arxiv.org/abs/1707.07012 computational cost Accuracy(precision@1) accuracy AutoML outperforms handcrafted models
- 21. ML on Google Cloud Platform Application developers Machine Learning APIs Cloud AutoML Cloud ML Engine Data scientists & ML practitioners new
- 22. Cloud Machine Learning Engine cloud.google.com/ml-engine Fully Managed platform for machine learning Distributed Training with GPUs & TPUs Fast & Scalable online/batch prediction
- 23. About 10 lines of Keras to train a neural network https://www.tensorflow.org/tutorials/
- 24. 2 Speed UP Your AI Lifecycle
- 25. Up to 180 teraflops per device 64 GB high-bandwidth memory Real images/sec ~2650 Train to 76% top-1 accuracy on ImageNet for < $85 Cloud TPU
- 26. Cloud TPU v2 180 teraflops 64 GB HBM training and inference Cloud TPU v2 PodALPHA 11.5 petaflops 4 TB HBM 2-D toroidal mesh network training and inference Cloud TPU v3ALPHA 420 teraflops 128 GB HBM training and inference Google Cloud TPU horsepower On-prem GPUs depreciate at 50% pr. year. Use the cloud instead
- 27. Serverless analytics for complete ML lifecycle on Big Data Ingest Explore Prepare Pre- process Train Hypertune Test Predict (Online) Predict (Batch) ML activity GCP services Transfer Service GCS Pub/Sub BigQuery Dataprep Dataflow Dataproc BigQuery Dataprep Dataflow Dataproc BigQuery Data Machine Learning Engine Apps https://cloud.google.com/solutions/big-data/ 80-90% of the work Value
- 28. 3 Embed Your AI Everywhere
- 29. TensorFlow Lite TensorFlow TensorFlow.js Train and execute Models in the browser! Run inference on mobile and IoT devices Train and execute Models on a variety of devices Different scenarios for machine learning https://js.tensorflow.org/
- 31. Edge TPU module (launch October) Works with: SoM combining: Quad-core CPU Wifi Secure Element (Crypto chip) Edge TPU
- 32. Edge TPU development kit Available through ➤ System On Module (SOM) = Integrated Edge TPU + CPU + Crypto ➤ Edge TPU accelerator for edge systems that already have a host connectivity ➤ Dev board adds connectivity options to SOM
- 33. Demo at Next Camera input 1080P at 60 FPS Multiple AI models run concurrently on Edge TPU No Edge TPU 3-4 FPS Edge TPU 25 FPS https://youtu.be/-T9MNR-Bl8I?t=1697
- 34. Image analysis of 200 HD images of glass panels in 0.8 seconds Edge TPU built into camera From 60% to 99.9 % accuracy https://www.facebook.com/gcp/videos/lg-cns/1026946984153368/
- 35. ML lifecycle with Cloud IoT Edge Device data Aggregated data Build or bring your own model Trained ML model Data Deploy trained ML model Build and train ML model in the cloudCloud IoT Edge Real-time analytics & ML Operational data
- 36. Democratize Your AI Speed up Your AI lifecycle 1 2 3 Embed Your AI everywhere Henrik Hammer Eliassen Customer Engineer Google Cloud Three directions to make AI more useful