Quoc Le, Software Engineer, Google

on YouTube ↗26m

" summary

Code and models are released at tensorflow.org, with full source so researchers can reproduce results. A sequence-to-sequence recurrent model handles translation, conversation modelling and speech recognition and matches prior best baselines. Large-scale training runs on thousands of machines (example: 16,000 CPUs) to train networks with about a billion synapses. Empirical gains include image error dropping from ~28% to ~5% and speech word-error falling from ~30% to ~8%, demonstrating practical, production-grade improvement.

AI-written summary · about

" chapters (16)

1. 0:00Introduction
2. 1:20Motivation
3. 3:28Deep Neural Networks
4. 7:00Parameter Server
5. 8:04Shallow ML
6. 8:59YouTube Deep Learning
7. 9:57The watershed moment
8. 10:14Machine learning can invent features
9. 10:54Googles ML platform
10. 11:19Image recognition
11. 12:09Voice search
12. 13:22Text understanding
13. 16:02Training
14. 18:51Meaning of life
15. 23:01Tensorflow
16. 24:36Questions

" related talks

• 24mQuoc Le, Software Engineer, GoogleQuoc Le · MLconf SF 2014
• 07mMachine learning at GoogleRob Craft · Strata + Hadoop World 2017
• 26mUsing deep learning in natural language processingRob Romijnders · PyData Amsterdam 2017
• 59mMachine Intelligence at Google ScaleGuillaume Laforge · Voxxed Days Singapore 2017
• 09mInterview with Martin GörnerMartin Gorner · Devoxx Belgium 2017
• 27mKeynote - Google Cloud, Machine LearningKarthik Padmanabhan · GDG DevFest Ahmedabad 2018
• 13mTaking Machine Learning to the Next LevelGiovanni Galloro · Codemotion Rome 2018
• 04mState-of-the-art Machine Learning research with Google toolsLaurence Moroney · Google's Machine Learning Virtual Community Day 2021