Talk:One-shot learning (computer vision)

Article too narrow - needs history

I really oppose to the restricted representation in the current article. One-shot learning (single-shot learning) is a much older concept than is presented here. It is long known in biology and experimental psychology since the beginning of the 1980-ies. — Preceding unsigned comment added by 129.125.178.72 (talk) 13:19, 16 August 2016 (UTC)

This is because the article seems to be text copied from one or two research papers. It includes textual references to images which are not even included. It really needs a rewrite from someone who has a broader knowledge in this area. 213.16.80.50 (talk) 09:16, 15 November 2016 (UTC)

+1 vote for that. Needs a complete re-write.

2601:199:C301:D4FC:D531:A618:996C:9484 (talk) 03:12, 15 March 2019 (UTC)ML

The idea of one shot learning is also not limited to object categorization. DavidCJames (talk) 18:12, 7 February 2023 (UTC)

Redeveloping this article

Hello everyone, I am mostly inactive, but I came across this article and think that it could use some work to improve its encyclopedic tone, and I am currently in the process of developing some thoughts on it in User:Gluons12/One-shot learning. Right now I'm still quite a while from actually doing a significant rewrite, as I expect I'll need to do some research myself before I can tackle it, although I do have some background in computer science. But I wanted to drop a notice of what I think needs adjustment in the article in case of objections. The overarching goal is to address the banner to create a more encyclopedic tone. 1) As mentioned in the above section, trying to write a more accessible history of the topic, if sources are available. 2) Reduce reliance on research papers and seek less technical sources for overview purposes. 3) Update technical sources to reflect the current state of the research. (survey papers will likely be useful here) 4) Reduce/eliminate the mathematical details which are very heavy compared to comparable articles (e.g. machine learning, convolutional neural network, generative adversarial network, etc.) Since actually implementing this will likely involve removing substantial amounts of current content, I wanted to advertise first. Once again, as my activity will inevitably vary with my level of business in my real life, I can't say with certainty when I'll get around to the actual rewrite. Gluons12 ^t|c 20:39, 24 April 2021 (UTC).

Links beyond robotics

Hi all - I came across this article, and decided to contribute to this talk page, because I am an educator at an archives who often delivers "one-shot instruction" workshops. This is a common term within the library and archives profession. I was surprised to find that this article does not reflect this usage of "one-shot" at all, nor the lengthy history of scholarship within library studies and among librarians on the topic. I don't know enough about robotics to undertake the full revision of the article recommended above (+1, sounds necessary!), but wanted to add that this revision should include mention of one-shot instruction as it is a huge field of practice and study. Suggest Buchanan and McDonough's "The One-Shot Library Instruction Survival Guide" (ALA Editions, 2021) as a starting point. Alison.d.little (talk) 20:12, 28 June 2021 (UTC)

هتل رزرومن کجاست

فرداهوا چطوره ؟ 91.133.208.231 (talk) 21:26, 14 January 2022 (UTC)

Renaming to match scope

With the exception of the recently added "In natural language processing" section, this article is entirely about one-shot learning for image classification. But because it has the generic title One-shot learning, it has a lot of inappropriate incoming links which are referring to OSL for other problems. We could try to expand the scope of this article to cover the paradigm more broadly, but there's already way more detail about the computer vision application than would be appropriate for a broad-concept article on one-shot learning, and I wouldn't want to just delete all that content, so I'm going to instead:

• Rename this article to something like One-shot learning in computer vision (if someone else can think of a better title, feel free to move it again). I will not leave behind a redirect.
• Go through the incoming wikilinks, and update any that actually should be pointing to this article.
• Merge the "In natural language processing" to some more appropriate destination (maybe prompt engineering)

We should definitely have a broad article about one-shot learning, but the question of scope is tricky. Should it also incorporate zero-shot and few-shot learning? It certainly doesn't help that these terms have been used with different meanings in different fields and at different times. We do have Zero-shot learning, which could be a good target for now. I need to read through it more carefully to try to understand its current intended scope. Colin M (talk) 15:29, 14 March 2023 (UTC)

In addition to zero-shot learning, we also have Explanation-based learning. I am begging academics to please stop using a bunch of different names for the same thing. :( Colin M (talk) 15:45, 14 March 2023 (UTC)

@Colin M: I support this move for the reasons you listed, but I also do believe that one/few-shot learning in the context of NLP warrants its own article. I propose starting the article few-shot learning and having it address few-shot learning in NLP, maybe with a subsection about one-shot (that may not really be necessary, but we'll see). If there are no objections, I'll start the article later today. –⁠ ⁠Popo ⁠Dameron ⁠talk 17:34, 14 March 2023 (UTC)

@PopoDameron: that sounds great! A few thoughts/questions:

• Few-shot and one-shot are similar enough that I would definitely cover them in the same article. One-shot is just a special case of few-shot with n=1, and I don't think it's conceptually different from n=2 or 3 or whatever.
• If you want to restrict the scope of the article to NLP (which I support - there's definitely enough sources and content to support a standalone article with that scope), I would give it a title that reflects that scope, so it doesn't attract bad incoming links (since it does seem like the term "few-shot learning" is also used in other domains including computer vision). Few-shot learning in natural language processing? Or maybe just Few-shot prompting?
• Worth thinking about how the scope can be disentangled from Prompt engineering - that article seems a bit muddled at the moment, but it definitely has some overlap with this topic.

Colin M (talk) 17:43, 14 March 2023 (UTC)

@Colin M: Sounds good! I think I'll go with Few-shot learning in natural language processing and we can probably just disambiguate on Few-shot learning and One-shot learning. –⁠ ⁠Popo ⁠Dameron ⁠talk 17:49, 14 March 2023 (UTC)

Oh, and one other consideration is that, maddeningly, the term "zero-shot" has, historically, sometimes been used with a sense of what we would now more commonly call few-shot or one-shot. The "Language Models are Few-Shot Learners" paper mentions this on page 2 in reference to the earlier GPT-2 paper. Given that, at the very least we should probably provide some disambiguation from Zero-shot learning. Colin M (talk) 17:47, 14 March 2023 (UTC)

Ah yes, of course.

As part of this investigation, we also clarify and systematize the approach introduced in [RWC+19]. While [RWC+19] describe their work as “zero-shot task transfer” they sometimes provide examples of the relevant task in the context. Due to the use of what are effectively training examples, these cases are better described as “one-shot” or “few-shot” transfer

–⁠ ⁠Popo ⁠Dameron ⁠talk 17:51, 14 March 2023 (UTC)