Look at the errors
Error analysis means manually inspecting a sample of wrong predictions and categorizing them. A single number hides where the model fails. Reading fifty mistakes tells you what to fix next.
- Pull a random sample of misclassified or high error cases.
- Tag each with a likely cause such as blurry image, rare class, or label noise.
- Count the tags to rank causes by frequency.
Prioritize by impact
The biggest tag is usually the best place to spend effort. If thirty percent of errors come from mislabeled data, cleaning labels beats tuning architecture.
- Estimate the ceiling gain if a category were solved.
- Weigh that against the effort to fix it.
- Fix the cheap high impact category first.
The loop
This turns a flat metric into a prioritized to do list.
Key idea
Error analysis groups individual mistakes into causes, ranks those causes by potential gain, and points you at the fix with the highest return on effort.