The scientific method

The scientific method is not one fixed recipe, but most research follows the same loop.

1. Observe. Notice something: a pattern, a problem, a result that does not fit.
2. Ask a question. "Why does this model do worse on photos taken at night?"
3. Form a hypothesis. A possible answer you can test: "The training data has very few night photos."
4. Predict. If the hypothesis is true, what should happen? "Adding night photos to training should improve night accuracy more than adding the same number of day photos."
5. Test. Run the experiment or collect the data.
6. Analyse. Compare what happened with what you predicted.
7. Share. Write it up so others can check it.

Then the loop starts again. The result raises new questions.

Testable means "could be wrong"

The philosopher Karl Popper argued that a scientific claim must be falsifiable: there must be some result that would show it is wrong. "Night photos are under-represented" is falsifiable: you can count them. "The model has a hidden essence that dislikes darkness" is not: no test could ever show it false.

This is why a good prediction is specific. "Accuracy will change" can never fail. "Night accuracy will rise by at least 5 points" can.

Results support, they do not prove

One experiment that matches your prediction supports your hypothesis. It does not prove it. Another explanation might fit the same result. Maybe the night photos you added were also sharper, and sharpness was the real cause. Good researchers look for other explanations and try to rule them out.

RememberWrite your prediction down before you run the test. It stops you from quietly changing the question to fit the answer.

ExampleYou predict that a larger learning rate will make training faster. You test it and training is faster, but the final accuracy is worse. That is a real result: report both parts, not just the part that matches your prediction.