TensorTau PathToAGI Doing Machine Learning Honestly Level 05

Level 05 · Doing Machine Learning Honestly · Chapter 075

How to Avoid Cherry-Picking

If we try enough seeds, metrics, datasets and plots, can we accidentally “discover” the result we hoped to see?

1. Researcher degrees of freedom

Suppose we try 20 hyperparameter sets, 10 seeds, 5 metrics and 4 data subsets, then publish only the nicest number. Even without deliberate dishonesty, the workflow selected for luck.

Cherry-picking is not just lying. It is selective visibility.

2. Decide important rules before seeing the answer

Whenever practical, predefine primary metric, evaluation split, stopping rules, baseline, comparison method and exclusion criteria.

Exploratory work is valuable—but label it as exploration and validate resulting hypotheses on new evidence.

3. Report the distribution

Show all seeds or a principled summary, not only the best one. Report failed or negative ablations when they affect interpretation. Separate model-selection data from final evaluation.

4. Keep an experiment ledger

time | hypothesis | commit | data | config | seed | metric | outcome

This record prevents unconscious rewriting of history after seeing results.

5. The Level 5 checklist

Before believing “our model improved” ask:

  • Was evaluation data truly unseen?
  • Was there leakage?
  • Is the metric appropriate?
  • Did it beat a meaningful baseline?
  • Is the result stable across runs?
  • Is uncertainty reported?
  • Were ablations controlled?
  • Were many comparisons hidden?
  • Would the conclusion survive a fresh dataset?

What to Remember

Honest ML is not only about using correct equations. It is about designing a process in which we would have been able to discover that our favourite idea did not work.

This closes Level 5.

Next → Meet the Smallest Neural Network: now that we know how to evaluate learning honestly, we can build the machinery behind deep learning one number at a time.

Six ways to work the chapter, climbing from explaining it in your own words to investigating it like a researcher. Reading is not enough — this is where the chapter lands.

  1. Explain: Distinguish exploration from confirmatory evaluation.
  2. Calculate: Count the number of potential reported results from 10 seeds × 8 configs × 4 metrics.
  3. Predict: Why does repeatedly checking the test set turn it into a validation set?
  4. Break: Simulate many equal models and publish only the best score; quantify selection bias.
  5. Build: Create an experiment ledger template and use it for a real mini-study.
  6. Investigate: Reproduce one earlier result using a prewritten protocol and fresh seeds/split; report whether the original conclusion survives.