Conference Presentation, Keynote, Fireside Chat
Jeff Dean’s Lecture for YC AI
- The Google Brain Team plans to continue open-sourcing TensorFlow to foster a global community, aiming for a system that supports flexible research, rapid experimental turnaround (reducing timelines from weeks to minutes or hours), and broad deployment across data centers, mobile devices (iOS, Android), Raspberry Pis, and custom accelerators.
- Infrastructure scaling goals include achieving nearly linear speedups for image models on up to eight GPUs and close-to-linear scaling for 64 GPUs, with a commitment to release 1,000 free TPU devices and Cloud TPUs (featuring 180 teraflop TPU v2 VMs) later in the year to accelerate training and inference.
- Specific application predictions include optical character recognition for Street View images, solar energy potential estimation for rooftops, and a diabetic retinopathy classifier expected to match or slightly exceed the performance of eight U.S. board-certified ophthalmologists, with clinical trials currently underway in India.
- Robotics strategies involve using simulators to generate training data, with a neural net-based simulator projected to be 300,000 times faster than real simulators to process 100 million scenarios, while also transferring learned behaviors from video and simulation to real-world robotic arms for grasping and manipulation in messy environments.
- Natural language processing ambitions include scaling translation models to utilize 100x to 1,000x more training data than academic datasets, employing deep LSTM stacks and attention modules to achieve human-level or superior quality for specific pairs like English to Japanese, and automating architecture search to find optimal structures for Cifar-10 and language modeling tasks.
- Automated machine learning initiatives focus on "learn-to-learn" approaches, such as generating models without human intervention, creating learned optimizers that outperform Adam, and training model-generator systems to produce architectures in hours, aiming to run 12,000 experiments over a weekend.
- Future system architectures anticipate the development of a single giant model capable of thousands to 10,000 tasks via sparse activation (100 billion to 1 trillion parameters), combining memory networks for working memory, and integrating learned cells that outperform LSTMs on medical record prediction tasks.
- Cloud services are planned to expand with higher-level APIs and pre-trained models, including a Vision API for reading text and identifying objects in photos, and a Translation API, alongside optimizations that place TensorFlow operations on devices 20% faster than human-designed configurations.