The Tuning Fallacy

The Tuning Fallacy operates as follows:

• A breadboard system is trained on one set of data and achieves 90% accuracy on that data. This is taken as a ``proof of concept'' .
• However, the trained system only gets 70% accuracy on a new data set.
• If trained on the new data set, the system reaches 90% on that data, but only gets 70% on the original data: the training set phenomenon or overfitting.
• The goal becomes to find a setting of the weights that will recognize all the exemplars simultaneously.

The Fallacy: There is no such setting. But one can spend years and millions of dollars looking for one.

The problem is the false alarm rate as the number of exemplars grows.

The ROC Curve[Receiver Operating Characteristic] is often used to show true positive rate recall plotted against false positive rate.