2
$\begingroup$

The Wikipedia Article on one-shot learning described it as

Whereas most machine learning based object categorization algorithms require training on hundreds or thousands of images and very large datasets, one-shot learning aims to learn information about object categories from one, or only a few, training images

I don't think this is a particularly good definition, since it is very computer vision focused, but maybe I am wrong.

I have a method, which I would like to describe as oneshot learning. But I am not sure if it is the the right term.

  • We are not doing computer vision at all.
  • We are learning feature vectors for classes -- think something like a bottle-knecked deep neural network (It is not that).
  • Other methods for learning these feature vectors use many (hundred, thousands) of labelled examples of each class -- sometimes after a unsupervised pretraining step, sometimes not.
  • We make use of particular structure in the problem/feature vector to instead only need a single labelled example of each class, after using a unsupervised pretraining step.
  • but we don't use any existing machine learning method to do so. Infact it is arguable if our method is machine learning at all. The unsupervised part certainly is, but the final step, with the label data ... well it doesn't have any direct optimization step which I have come to expect in a machine learning process (No EM, or Gradient descent). Effectively it is a informed kind of averaging, that works because of the structure of the representation/feature vector.

Is it still appropriate to call this one-shot learning?

$\endgroup$
3
$\begingroup$

The Wikipedia article you mention, One-shot learning, does seem specific to computer vision.

A broader term is semi-supervised learning, which may be more appropriate.

(I am not an expert in Machine Learning or its various sub-communities, but the above is consistent with the first pages of results on Google Scholar for these terms.)

$\endgroup$
  • $\begingroup$ Issue is that a typical Semisupervised learning approach tends to have a significant amount of labelled data. I want to emphasize how little labelled data we use. $\endgroup$ – Lyndon White Oct 19 '16 at 0:59
  • $\begingroup$ As I said, I am not really in these communities. The "one shot" term seems to fit, and is descriptive*, so I would be OK using it. You could look in journals/conferences where you would want to publish, to see what terms are most appropriate for your community. (*IMO at least as much as, say, "semi-unsupervised learning" would be.) $\endgroup$ – GeoMatt22 Oct 19 '16 at 1:10
  • $\begingroup$ I'm less concerned with familiarity with the term (as you say, it seems to fit and is descriptive), and more concerned with whether or not it suggests I am using one of a particular family of algorithms. (rather than solving a particular kind of problem) $\endgroup$ – Lyndon White Oct 19 '16 at 1:18
  • $\begingroup$ Google Scholar gives 642,000 results for "one shot learning", and that number, combined with a skim of a couple of pages of titles, would suggest to me that it is more likely about a class of problems. (You can skim the link to see if you feel the descriptor is appropriate for your case ... "de facto" usage, if you will.) $\endgroup$ – GeoMatt22 Oct 19 '16 at 1:34

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.