1
$\begingroup$

The Show, Attend and Tell paper describes a solution to image captioning.

I am wondering: Is this novel machine learning application and instances of (semi-)supervised or unsupervised learning (in the broadest sense)? And if not, why?

My confusion comes from the fact that for disconnect that I'm experiencing, when learning introductory machine learning concepts, where everything is categorized as a (semi-)supervised problem, where the output is either a vector or a label or unsupervised learning problem, where clustering is performed - and the above solution, whose output seems to be a much more complicated object, for example a word.

I guess one could encode words in vectors or labels drawn from a large pools of possible labels, but I'm very doubtful if this is actually how it is done.
The machine learning solutions from the textbooks, that I learned so far, only seem to apply for problems where the output is a number (such as predicting a the income in a country based on years of education and age) or a label (such as in the Iris dataset). But the linked problem seem of a very different category.

(I tried going over the details of the Show, Attend and Tell paper. While I do know a bit of stats, I lacked machine learning knowledge to be able to discern whether their is ultimately a supervised or unsupervised machine learning solution.)

$\endgroup$
  • $\begingroup$ Are all these novel machine learning applications instances of (semi-)supervised or unsupervised learning (in the broadest sense)? - no. As to why, you should ask separate questions (for example answer to image captioning is pretty complicated). $\endgroup$ – Jakub Bartczuk Apr 29 '18 at 15:24
  • $\begingroup$ @JakubBartczuk I'm not really interested in a fully explanatory why they are not (un)supervised learning example - but rather, under what classification they would fall in that case. If that is still too large for a single answer, just pick any of the mentioned example to give an example - I edited the question accordingly. $\endgroup$ – l7ll7 Apr 29 '18 at 16:30
  • $\begingroup$ @JakubBartczuk Edited yet again, now I'm solely focussing in the Show, Attend and Tell paper $\endgroup$ – l7ll7 Apr 30 '18 at 6:31
  • $\begingroup$ Please 1. edit the title to focus solely on this paper 2. give a full refrence to the paper? $\endgroup$ – Juho Kokkala May 1 '18 at 9:06
2
$\begingroup$

This paper is very much a supervised learning problem.

I guess one could encode words in vectors or labels drawn from a large pools of possible labels, but I'm very doubtful if this is actually how it is done.

You are right. Words are typically represented as a categorical variable. The vocabulary size for neural models is typically on the order of 10000 (as it was for this paper).

$\endgroup$

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.