A Technical Deep-dive into How Drive Uses ML to Boost Enterprise Productivity
summary
Drive uses a multilayer feed-forward neural network with three hidden layers of 180 ReLU units, a sigmoid output, and a ~14,000+ sparse inputs vector of file and user features to predict open probability. Training is supervised on tens of billions of training examples from Drive logs using TensorFlow on Google Cloud and BigQuery. Production relies on AB experiments, automated dashboards and strict launch criteria; Quick Access cuts finding time ~50%, hit rate rises ~11 percentage points and saves ~850 workweeks per day.
AI-written summary · about
chapters (10)
- 0:00Introduction to Drive ML
- 2:33Overhead in knowledge work
- 5:21Quick Access launch and function
- 13:38Priority homescreen and workspaces
- 17:36Building blocks of Drive intelligence
- 25:25ML infrastructure and training
- 30:22Experiment velocity culture
- 36:50UX improvements and hit rate
- 40:44Trust privacy and reliability
- 42:08Recap and final stats
