It appears to me that "feature extraction" and "feature learning" are equivalent concepts, however there are 2 separate wikipedia articles dedicated to them that are notably different. In particular, only in the Feature Learning article Neural Networks/Deep Learning are mentioned. However, it seems like they would be equally appropriate for either because autoencoders extract features from the raw (typically image) data and those now extracted features feed into the next layers.

So, what's the real difference between these terms?


Yes I think so. Just by looking at Feature Learning and Feature extraction you can see it's a different problem.

Feature extraction is just transforming your raw data into a sequence of feature vectors (e.g. a dataframe) that you can work on.

In feature learning, you don't know what feature you can extract from your data. In fact, you will probably apply machine learning techniques just to discover what are good features to extract from your dataset. Then you can extract them them apply machine learning to the extracted features. Deep learning techniques are one example of this.

In the word2vec toolkit, for instance, you extract vectors from documents which can't be easily interpreted by a human, you can't look at it and tell what features have been extracted at all. It's just a mass of vectors which, for some reason, give good empirical results.

| improve this answer | |
  • $\begingroup$ Thanks! I'm still not 100% convinced: "Feature extraction is a general term for methods of constructing combinations of the variables to get around these problems while still describing the data with sufficient accuracy." Now, of course "combinations of the variables" means new features, basically learned. I agree that this is a discussion about terminology, but I'm trying to learn what's what. $\endgroup$ – Tnatsissa H Craeser Nov 9 '15 at 23:41
  • 1
    $\begingroup$ Feature extraction is more applying some pre-defined rule/function on data and getting those features. Feature learning is more task-oriented, e.g. you learn features to recognize digits as your ultimate goal. Both are similar in a sense that they are just a mean to get a learning algorithm clues to do its predictions. $\endgroup$ – Vladislavs Dovgalecs Nov 10 '15 at 0:27

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.