4
$\begingroup$

I've recently read about concept learning in a machine learning class. They defined a concept as (translated from German):

Concept:

  • Describes a subset of objects or events defined on a bigger set.
  • Boolean function defined on bigger set

Given: Examples which are labeled as members or not-members

Searched: Automatically infer the definition of the underlying concept

Definition of concept learning: Infering a boolean-valued function from training data of its input and its output.

I would call this a 2-class classification problem. Is there a difference?

|improve this question
$\endgroup$
  • $\begingroup$ Nice question. Are there any online links, which can help us learn more about the concept of concept learning? $\endgroup$ – Dawny33♦ Dec 2 '15 at 7:11
2
$\begingroup$

The objective is rather different. For classification problems, we want to know: GIVEN X, WHAT WILL BE Y? For concept-learning problems, we are asking: WHAT KIND OF X WILL GIVE US Y?

It may seem similar, but for classification problems, we may use a model that is unable to generalize and answer "WHAT KIND OF X WILL GIVE US Y" until you plug X into the model (e.g. K-nearest neighbor).

On the other hand, concept-learning problems, the main issue is to generalize and answer what can be X specifically.

|improve this answer
$\endgroup$
  • $\begingroup$ Which algorithms would be used for concept learning? $\endgroup$ – Martin Thoma Nov 2 '15 at 5:54
  • $\begingroup$ This slide: ml.informatik.uni-freiburg.de/_media/documents/teaching/ss09/ml/… outlined a few algorithms for concept learning $\endgroup$ – Jacky Ma Nov 2 '15 at 6:04
  • $\begingroup$ So with techniques like the ones used in deep dream / the ones presented by Matt Zeiler in Visualizing and Understanding Deep Neural Networks one can say that the CNN is learning the concept of "cats" and can of course be used to classify images as "cats"? $\endgroup$ – Martin Thoma Jan 7 '16 at 10:53
  • $\begingroup$ It still seems to me that any learning algorithm for classification can be used to generate instances of the class. Hence all classification algorithms would also be concept learning algorithms? $\endgroup$ – Martin Thoma Jan 7 '16 at 10:54
  • $\begingroup$ No. classification only has a model p(y|x), so can't generate x's given y -- p(x|y). You could do that with a generative model only, which learns p(x,y), like naive bayes for example. Maybe this is what concept learning is in modern terminology? $\endgroup$ – samlaf Aug 9 at 15:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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