Corinna Cortes, Head of Research, Google, NYC 2017
summary
Describes concrete methods to make neural nets production-ready: AdaNet jointly learns architecture and weights with provable generalization bounds and empirical gains; outputs get constrained with finite-state transducers so entities and numbers stop hallucinating and entity accuracy jumps from poor to near-perfect for measurements; large softmax layers speed up by replacing millions of weight vectors with a 256‑centroid weight quantization plus approximate nearest‑neighbor search, cutting inference work by orders of magnitude while keeping accuracy.
AI-written summary · about
chapters (26)
- 0:00Intro
- 0:45Barriers to full networks
- 3:00Googles mission statement
- 4:13Google Translate
- 4:41Smart Box
- 5:33LSTs
- 6:39Music Projects
- 7:10Smart Replies
- 10:26Regular Expressions
- 10:58Text Summarization
- 12:47Channel Model
- 13:59Errors
- 14:46Numbers
- 16:21Grammar
- 18:52Neural Networks
- 19:46Classification
- 20:19Speed
- 22:08Reform
- 24:12Scaling
- 25:59Deployment
- 26:35Finding good network architecture
- 28:35Learning network architecture
- 34:49Reinforcement Learning
- 35:58Tables and Accuracy
- 37:04Google Paper Cuts
- 37:21Conclusion
related talks
48mKeynote – Sameer Singh
53mML BuzzWords demystified
28mDemystifying Deep Neural Nets
09mLightning Talk: Intellectual Cognitive Development in Artificial Intelligence
41mDeep Learning to Solve Challenging Problems
01h 14mRecent Developments in Over-parametrized Neural Networks, Part I
37mCorinna Cortes, Head of Research, Google
39mA Fireside Chat with Turing Award Winner Geoffrey Hinton, Pioneer of Deep Learning