Imagine you want to diagnose a skin lesion.
You look at it under a dermatoscope, assess it based on the criteria and cases you’ve seen before, and make a diagnosis.
If you had no point of comparison, this would be impossible.
But, by referring back to fundamental criteria, you can reach the right conclusion.
This is the essence of supervised learning.
We show a model lots of inputs (e.g. a skin lesion), and show it the right output (e.g. “this is benign”), and let it develop pattern recognition.
With enough experience, we can remove the output, and let the model infer the right answer based on past experience.
Supervised learning can use classification when there are a finite number of outputs (e.g. is this lesion benign or malignant?), or regression when there are infinite outputs (e.g. what’s this patient’s five-year cardiovascular risk?).
With more data, the model gets better and better at figuring out the right answer.
On the other hand, in unsupervised learning, we provide no “right answers”.
Instead, we feed a model a bunch of data, and ask it to find its own pattern.
Often, computer find patterns that we’d never notice.
For instance, we now have models that can detect BMI, diabetes, hypertension, smoking history, sex, kidney, liver, and gall bladder disease, and risk of heart attack, stroke, Alzheimer’s and Parkinson’s — all from a retinal image.
Incredible, right?
Let’s explore a simple kind of supervised learning in more detail.