# If machine learning, in general, is about ‘learning’ patterns, is it correct to say its branches differ based on the type of pattern being learnt?

If machine learning, in general, is about ‘learning’ patterns, is it correct to say its branches differ based on the type of pattern being learnt?

I.e. in supervised learning the pattern is a target function, in unsupervised learning, it’s some structure like clusters, in RL I guess it’s a reward-maximising action given a state.

• I'm not sure I would put it that way. E.g., this to me does not fit this: the internal representations in neural networks, for which it often turns out that representations that are created by training using a self-supervised/unsupervised approach (e.g. using in-painting for images, denoising autoencoders for tabular data, or the BERT-style "fill-in-the-blanks" pre-training objective for language) are pretty good for classifcation if you just slap a logistic-regression style final layer on top. If you think that's not "patterns", then perhaps the term/the statement is too vague to be useful? – Björn Jun 16 at 10:38
• It is not about learning patterns – Aksakal Jun 16 at 14:01

Operational definition of Machine learning is probably best defined by Tom Mitchell's book Machine Learning, Machine Learning is the study of computer algorithms that improve automatically through experience. A bit more formally,

A computer program is said to learn from experience E
with respect to some class of tasks T and performance
measure P, if its performance at tasks in T,
as measured by P, improves with experience E.


This implies being able to generate learning curves and supervised learning only.

Clustering being unsupervised learning is ill-defined, because unsupervised learning initially was referring to Hebbian Learning. Reinforcement learning is really a different beast compare to pervious two and originates from study of how animals learn, see Watkins story on how he start working on reinforcement learning.

It is true that different branches of Machine learning indeed have different understanding of what constitute by learning. They don't agree on what is a pattern either. Instead of saying its branches differ based on the type of pattern being learnt, if we say its branches differ based on understanding of what is learning would probably make more justice.

I'm not sure if there is strict definition of the branches of machine learning, (e.g. is deep learning a branch) but if we consider [semi-]supervised, unsupervised and reinforcement learning as major approaches, I think the most important distinction is the problem formulation, i.e. what is to be learnt. Patterns learnt might be the same, similar or different based on this definition.