# What's the threshold to call something 'machine learning'?

For example, if I use some iterative solvers to find a solution to a non-linear least squares problem, is that already considered machine learning?

• This question seems to be a duplicate of ai.stackexchange.com/q/12558/2444. Can you clarify how your question is different from that other one?
– nbro
Sep 24, 2020 at 13:56
• @nbro: I think your potential duplicate asks specifically for a definition, whilst this question is asking for a way of classifying a problem as ML or not, when it might seem similar. This one goes into detail where the linked one stops. Sep 24, 2020 at 14:08
• @NeilSlater Well, a problem is just a problem. You can apply different techniques to solve it, so asking if a problem is machine learning does not make much sense, to be honest, but I didn't think much about it too. Asking which kind of problems can be solved with machine learning techniques makes more sense. This question seems to be about "what kind of techniques are considered machine learning techniques", which, I agree with you, is slightly different from the more general question "what is machine learning".
– nbro
Sep 24, 2020 at 14:11

## 1 Answer

T. Mitchell defines machine learning in "Machine Learning" book as

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

Hence, based on the above definition, we can't say a machine learning method to every iterative method. In your specific example, it is just a non-linear solver such as the Newton method to finding roots.

However, you should notice that a non-specific machine learning method can be used in the learning process. For example, you might need some numerical methods to compute the measure $$P$$‌ (in the above definition). But, we can't say that the specified method is a machine learning method.

• but wouldn't then for example an Autoencoder also not fall into the class of machine learning? Sep 24, 2020 at 19:41
• @user1282931 Sorry, but It's not a reasonable conclusion. Because, in autoencoder, the collected data for auto-encoding is E, and we have a loss function P (as a performance) as well. Hence, it is a machine learning method.
– OmG
Sep 24, 2020 at 19:45