Full Stack Deep Learning

1. Dream vs. Reality

Dream: build an optimal machine learning prediction system

Reality

• Aggregate, process, clean, label, and version the data
• Write and debug model code
• Provision compute
• Run many experiments and review the results
• Deploy model
• Monitor the predictions and close data flywheel loop

ex) Tesla’s self-driving system

• ML code portion is a lot smaller than the infrastructure

  

2. Three Buckets of Tooling Landscape

• data
• training/ evaluation
• deployment

3. Software Engineering