0
$\begingroup$

Is it even logically possible to have a type of machine learning that doesn't fall into those three paradigms?

Supervised: a dataset of inputs and outputs are fed to an algorithm which learns a function to generate outputs from inputs

Unsupervised: a dataset of only inputs are fed into an algorithm and the algorithm produces label outputs that can be used to make predictions from new data

Reinforcement: a dataset of only inputs are fed to an algorithm, which eventually learns to generate outputs that maximize a reward function

Isn't semisupervised ultimately just supervised with some output data missing? Is "punishment" learning viable? Would Hebbian learning count as something different? Most existing artificial neural net structures don't operate with the "fire together wire together" logic.

$\endgroup$
10
  • 1
    $\begingroup$ I’m not sure that I agree with your description of reinforcement learning. I think of RL as creating inputs (policy) that reliably give desirable outcomes. $\endgroup$
    – Dave
    Commented Nov 10, 2023 at 20:00
  • $\begingroup$ Isn't "desirability for a specific outcome" just another way of saying "value of the reward function for a specific outcome"? Also can you point me in the right direction to read about machine learning algorithms that don't fit under the main three or semisupervised? It's such a new field and 3 is such a small number. There has to be more. $\endgroup$ Commented Nov 10, 2023 at 21:07
  • 1
    $\begingroup$ Accurate predictions of the outcome matter in reinforcement learning, but the goal is to develop a strategy for what to do under various conditions to get desirable outcomes. Think of Go. The goal is to develop a strategy (“policy”) for what moves to make when the pieces are in particular arrangements. $\endgroup$
    – Dave
    Commented Nov 10, 2023 at 21:11
  • 1
    $\begingroup$ The question you ask would benefit more if you look at which types of problems each of the major learning categories solve. SL solves prediction problems ("learn to predict"), UL attempts to model/uncover the underlying distribution in a dataset, and RL solves optimization problems ("learn to optimize"). Please note the "LEARN TO" term to differentiate with other methods that do not use learning approaches. From these do you see some other problem formulations that do not fall into one of these categories? $\endgroup$ Commented Nov 12, 2023 at 6:13
  • 1
    $\begingroup$ If you examine statistical learning theory, which is used for prediction problems, you'll notice a generic formalization from which many other methods are derived. As a theorist, the question to consider is whether there exist any types of problems that do not fit into these three main categories (supervised, unsupervised, and reinforcement learning). You need to look at the most fundamental equations, understand what they do and move on from there. $\endgroup$ Commented Nov 13, 2023 at 3:09

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.