0
$\begingroup$

Maybe the question is too theoretical or even philosophical, maybe it's even the wrong SE-community.

I am wondering how I would call a model which is no longer trained/maintained with new data. Do I still call it Machine Learning? Since it does not learn any more.

Based on the following definition of ML:
“Machine learning is the study of computer algorithms that improve automatically through experience.” by Tom M. Mitchell
I would argue, since there is no automatic improvement anymore its not a machine learning algorithm anymore.

How do I call it when I used ML algorithms to train or fit a model to a certain point and then stop to improve the model?

$\endgroup$
4
  • 1
    $\begingroup$ I think that definition is a little liberal with the use of "automatically". These algorithms are capable of learning relationships without the user explicitly telling them the relationship. Once training has completed, the model is capable of incorporating more training data, but that comes at the users discretion. An algorithm in production is still machine learning, it just isn't learning at that moment. $\endgroup$ Nov 15, 2018 at 19:13
  • 1
    $\begingroup$ "Improve automatically through experience" refers to training. The Machine Learning part is about the computational methods used for training and validating a model. What you do with it afterwards is up to you. $\endgroup$
    – Digio
    Nov 15, 2018 at 19:20
  • 1
    $\begingroup$ If you want to read the name literally, algorithms are not machines, so machine learning does not exist... "Machine leaning is a study..." it's the name of whole field. Machine learning algorithm is an algorithm capable of learning, nothing guarantees that it will learn anything. $\endgroup$
    – Tim
    Nov 15, 2018 at 19:32
  • 1
    $\begingroup$ "How do I call it when I used ML algorithms to train or fit a model to a certain point and then stop to improve the model?. You call it a classifier or a regression function, depending on your context. Those are the end products of supervised machine learning algorithms. $\endgroup$
    – Zen
    Nov 17, 2018 at 4:36

2 Answers 2

2
$\begingroup$

It sounds like there is a mix up of the definition of online learning with the more general definition of machine learning. Online learning is an that area concerns itself with how to update a model when you don't necessarily receive all the data upfront, but rather want to update your model as data becomes available. But this is a subfield of machine learning, rather than a requirement.

I presume Tom Mitchell simply means that as the sample of data you use gets larger, your machine learning model should be able to predict better; i.e., a model built from 1,000 samples should predict better than a model built from 100 samples.

$\endgroup$
2
  • 1
    $\begingroup$ Thanks, your answer points into the right direction, I was thinking of a e.g. linear regression model. Of course a model based on 1000 data points is likely to predict “better” than a model fitted through 100 points. But when I stop to train after 1000 data points shouldn’t it be called machine learnt model? $\endgroup$
    – rul30
    Nov 15, 2018 at 20:21
  • 1
    $\begingroup$ @rul30: If you'd like to say "the machine has stopped learning", I think that's fine by all. But generally, "machine learning" refers to the field of study, not the actual action of learning from data. $\endgroup$
    – Cliff AB
    Nov 15, 2018 at 21:29
1
$\begingroup$

Machine learning algorithm is the full concept. These are deployed on machines that do not work when they are turned off, they "learn" only when deployed as learning algorithms, and it is an analogy to call what they do learning. However, that is no worse than normal language bending, e.g., calling a representative government, like France or Germany, a "democracy."

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.