In med school, we're taught to keep a broad differential.
We're told that, while common things happening commonly, we should never allow ourselves to get tunnel vision.
The exact same principle applies in machine learning. Which leads us to bias-variance tradeoff.
For instance, imagine an overly simplistic model. This is the med student who skimmed the textbook chapter, but didn't bother with the details.
Such a model has "high bias", meaning it's too simple to pick up nuances in data, leading to "underfitting".
But the inverse also applies.
If a model gets too accustomed to certain data, it may lose its ability to generalise, leading to "high-variance" models and "overfitting". This is the kid who always brings up semi-relevant zebra diagnoses just to remind everyone how smart he is.
The sweet spot, as you'd expect, lies in the middle.
And to better understand why this is, we need to understand the role of machine learning algorithms.