The Story of End to End Models in Deep Learning
summary
Baidu’s Deep Speech work claims that sufficiently large end-to-end nets beat hand‑engineered pipelines by directly mapping audio to text using models trained on 12,000 hours and then scaling to tens of millions of examples, needing massive training compute about 20 exaFLOPS over weeks, and achieving roughly 95% speech accuracy for many cases. Practical hurdles are data scarcity for robotics and dialogue, huge training and low‑latency deployment costs, and solutions offered include simulator‑generated data and batch dispatch on GPUs for serving.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 0:55Not EndtoEnd Models
- 2:22EndtoEnd System
- 3:15Object Recognition
- 4:31Why EndtoEnd Models
- 6:55Easier to Obtain Data
- 8:30Easier on Practitioners
- 9:17Crosstalk
- 10:21The flywheel
- 11:43Best example
- 12:58Challenges
- 14:26Compute
- 15:10Second Arrow
- 17:35Whats next
- 20:47Whats not coming
- 22:10Challenges in the future
related talks
42mdata.bythebay.io: Shubho Sengupta,
45mThe Deep Learning Revolution in Automatic Speech Recognition
01h 58mWhat If Your Backend Could Talk? Building Real-Time AI Voice Agents
43mWhy you can't afford to ignore deep learning
54mThe Future of AI: Building the Most Impactful Technology Together
25mEngineering voice agents: Latency, quality, and scale
13mAI: Big Expectations
25mFast deep recurrent net training