Tutorial, Conference Presentation, Fireside Chat
TensorFlow Tutorial (Sherry Moore, Google Brain)
Lex FridmanSherry Moore, Hugo Larochelle, Andrej Karpathy, Richard Socher, Ruslan Salakhutdinov, Andrew Ng, John Schulman, Pascal Lamblin, Adam Coates, Alex Wiltschko, Quoc Le, Yoshua Bengio, Shubho Sengupta, lexfridman, Zach, Pichin Lo
- Attendees will acquire tools to build applications including image recognition, color training, art generation, and music creation by the end of the tutorial.
- The tutorial aims to enable users to write, save, and immediately productize code without generating disposable prototyping scripts, aligning with the practice of researchers moving published prototyping code into production.
- TensorFlow's asynchronous data flow infrastructure is positioned as suitable for any application where computation triggers upon data readiness.
- Specific use cases for embedded deployment include running security systems on Raspberry Pi capable of capturing high-resolution images upon motion detection.
- Mobile response statistics via Smart Reply are projected to reach approximately 80% at the time of the talk, an increase from 10% recorded in February.
- The second lab curriculum covers checkpoint saving and loading, network evaluation, and the use of placeholders as critical infrastructure.
- Large-scale training workflows are described as involving the execution of numerous jobs, with performance monitored via morning loss plot reviews.
- Training times for the Inception model averaged six days on a single machine and reduced to two and a half days when utilizing 50 replicas.
- Windows support is anticipated to become available at some point in November, contingent upon Bazel reaching the necessary roadmap milestones.
- Integration with open-source frameworks such as MySource and HDFS for distributed training is in progress, though no specific completion date can be guaranteed.
- Java front-end support for Android integration is under consideration, with requests directed to specific team members for evaluation.
- Computing complex models like Inception on mobile devices is predicted to be prohibitively heavy for the hardware due to convolution and backpropagation demands.
- The team must consult with product leadership before committing to a timeline for supporting ARM processors on embedded boards like the TX1.