<aside> 💡 You are starting to work on a new problem and are in the process of constructing your model. What should you focus on initially? → Make sure your model can overfit!!!
</aside>
fastai-Tutorial about overfitting: Overfitting is harder to recognize because the predictions are very close to the data. This is not the case for underfitting, where the predictions are far away from the observations. Therefore we need a validation set. The metrics is measured only on the validation set.
The One Cycle Policy (fastai) will allow you to perfectly control overfitting.
https://twitter.com/radekosmulski/status/1574413519235084289