The Role of Software Developers in Machine Learning
summary
Claim: model training takes about 2% of project time, while data labeling runs $0.80 to several dollars per label and requires millions of images. Example build: an image-stitching pipeline that stores inputs and predictions in a searchable catalog of 4 million images, plus a playground and production model that auto-swap on better performance. Operational constraints: edge system must complete sense->act in 70 milliseconds. Common failures: compute, data-distribution, API/glue-code and test-environment mismatches, and missing drift detection.
AI-written summary · about
chapters (5)
- 0:00Intro
- 3:05Agenda
- 3:22What is the process for developing an ML project?
- 14:26ML infrastructure is complicated
- 19:45Where are the mismatches in the infrastructure?
related talks
04mManaging Machines
41mBeyond the Algorithm: What Makes Machine Learning Work?
30mChallenges of machine learning: The new ML lifecycle
04mThe age of machine learning
51mNot a magic : What to expect from Machine Learning projects
38mBuilding robust machine learning systems
36mMachine Learning Sucks
22mWhere to Start with Machine Learning