Amey-Thakur ML error analysis
The fastest way to improve a model is to look at what it gets wrong, one example at a time, until the failures organize into categories with different cures. An hour of reading errors routinely redirects a month of modeling.
Method
- Sample errors deliberately. From validation (never test; see train-test-discipline): the highest-confidence wrong predictions (model certain and mistaken: the most diagnostic), borderline cases near the threshold, and a random error sample as the control. 50-100 examples read by a human beats any aggregate metric for understanding.
- Tag each error with a cause hypothesis. Label wrong (annotation error), feature wrong (missing/stale/corrupted input: check the actual feature values; see feature-engineering), example ambiguous (humans would disagree), pattern unlearned (model lacks capacity or signal), out-of-distribution (nothing like it in training). Keep the taxonomy small; count the buckets when done.
- Let the counts pick the fix. Label noise dominant: clean labels or re-adjudicate guidelines (often the single highest-ROI fix). Feature bugs: fix pipelines (cheap, certain). Ambiguity: refine the target definition (see ml-problem-framing). Unlearned patterns with signal present: more features, more capacity, or more data of that kind: targeted collection beats bulk collection every time. OOD: expand training coverage or add an abstain path.
- Find the bad slices systematically. Metrics by segment (see model-evaluation), then automated slice discovery: train a shallow tree to predict "model was wrong" from features/metadata; its top splits name your weak cohorts in plain language. Compare error rates against the slice's base rate: a slice can be worse yet unimportant, or slightly worse and business-critical.
- Convert findings into regression tests. Each confirmed failure category becomes a named evaluation slice or behavioral test (perturbation and invariance checks) that future models must pass (see llm-eval-design for the generative analog); the analysis then compounds instead of evaporating after one fix cycle.
- Close the loop and re-analyze. After the chosen fix, rerun the same analysis: did the targeted bucket shrink, did another grow, did aggregate improvement come from where you intended? Error analysis is a loop, not a phase; each cycle's notes live with the runs (see experiment-tracking).
Boundaries
- Reading errors risks overfitting your judgment to validation; keep the final test set unseen, and confirm big decisions on fresh data.
- Feature-attribution tools (see shap-style explanations) complement but do not replace reading actual examples; attributions on a leaky or mislabeled dataset explain the artifact convincingly.
- Production error analysis inherits label latency and feedback-loop bias (see drift-monitoring); analyze recent labeled cohorts, not the convenient old ones.