In simple words: A useful model generalises to new cases.
Two common problems
Underfitting means the model misses patterns even in training data. Overfitting means it works very well on training data but poorly on new data. Good learning should carry over to unseen examples.
Watch validation results
If training loss falls while validation loss rises, overfitting may be starting. Early stopping pauses training near the best validation result. Better data, a simpler model, and dropout can also help.
See the flow
training result + validation result→diagnose fit
Simple example
A student memorises answers to 100 past questions but fails new questions with different wording. That resembles overfitting.
Remember: A useful model generalises to new cases.