I read in books, blogs, and articles that there are three learning paradigms: supervised, unsupervised, and reinforcement.

However, I have never found a proof that this list is exhaustive. Can it be proven that there cannot be other types of learning?

The notions 'supervised' and 'unsupervised' could be complementary to each other, like if $S$ is supervised and $\bar S$ is unsupervised, then $S\cap\bar S=\emptyset$ and $S\cup\bar S=U$ where $U$ is the set of all learning paradigms. However, this would be true only if $S$ and $\bar S$ were mutually exclusive and exhaustive. I don't believe that this is the case. First, there is reinforcement learning which is neither $S$ nor $\bar S$. Then, there is nothing that prevents learning to be not fully supervised or not fully unsupervised, e.g. labels can be not well defined, the teacher may teach only partially, but the system can still improve i.e. learn. And if there is reinforcement learning with is neither supervised nor unsupervised, can there be some other learning which is neither supervised, nor unsupervised, nor reinforcement?

Is it correct to say that in principle there can be many different learning paradigms?

  • $\begingroup$ Who says that? Who says there are only three paradigms? $\endgroup$ Commented Oct 12, 2023 at 19:09

1 Answer 1


You can formulate RL and unsupervised learning as "some sort of supervised learning"... in the case of UL you have that the target is an handcrafted task, it being similarity learning, next token prediction and what not, and in case of RL, the reward signal is by definition a form of supervision

The reason why we distinguish them it's because they aim at solving very different problems, but this is subjective

There are other types of learning, here are few of them:

  • semi-supervised learning
  • weakly-supervised learning
  • meta learning
  • self-supervised learning
  • transfer learning
  • anomaly detection

as per the others, you can lead these back to SL UL or RL, becuase at the end of the day, everything can be thought as an Energy model of some sort...

  • $\begingroup$ I'd would not say that anomaly detection is a type of learning. In fact, you can use unsupervised and supervised methods for anomaly detection. I'd rather say that it's a general task. $\endgroup$
    – nbro
    Commented Feb 3 at 16:34
  • $\begingroup$ @nbro anomaly detection includes dataset which contains both valid and unvalid points, thus it's not just density estimation (indeed, it's also known as outlier detection) $\endgroup$
    – Alberto
    Commented Feb 3 at 19:16

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .