Error analysis, data-centric iteration, baselines and knowing which experiment to run next.
8h · 8 lessons · 1 challenge
Before this: Optimizers & Training Dynamics
The highest-leverage skill in applied ML is choosing the next experiment, and it is almost entirely error analysis: look at the mistakes, categorize them, count them, and let the counts pick your work.
Covers baselines, human-level performance as a reference, train/dev/test discipline under distribution shift, and transfer and multi-task learning as tools rather than topics.
One week of GPU time and six candidate experiments. Choose, see the result, choose again. Ten weeks to hit the target metric.
Worth 500 XP.
