17

I’m working in the area of adversarial examples against neural networks. A standard research paper generally starts off an introduction with something along the lines of the below, but probably with a bit more guff:

Neural networks are increasingly used in safety and security critical applications, such as medical image recognition or facial recognition in surveillance footage. However, the networks are vulnerable to malicious attacks through the manipulation of input data. With adversarial examples...... etc.

Considering there are hundreds of papers which start with the same introduction, would it be acceptable to skip over writing my own guff? Possibly just reference existing papers and mention “existing uses and problems” (or something like that).

The scene has already been set by hundreds of existing papers, so why would I need to set the scene again in any of the papers that I create?

Or would something like this generally fail to pass peer review? Obviously there’s many different criteria for different journals/confs etc.

  • 1
    If we were being practical we would include hyperlinks to generally accepted definitions (like Wikipedia) instead of cluttering papers with half-baked explanations that are insufficient for new readers, and boring for those already in the know. But academia has never been the bastion of practicality. Hey, if your paper is accepted, you can go half way across the world to try and explain your paper in 10 minutes to 40 people. You have to accept the conventions that are there, you can't make news ones off your own bat. – Stumbler Aug 3 '19 at 7:51
  • 1
    Meanings of words change with time, would a reader in 50 gear time assume the same meaning as you do today? – Ian Aug 3 '19 at 9:30
  • Lots of good answers here, thanks all. I’m struggling to find just one to accept, as it’s quite a subjective question! – dijksterhuis Aug 5 '19 at 12:16
  • @dijksterhuisb please make the effort to accept one of the answers, otherwise the system will keep re-posting your question time after time after time, which is a really annoying part of the system. – Solar Mike Aug 16 '19 at 7:18
29

Your introduction should depend on the journal/conference you are submitting to as that will help define your audience. If you are submitting to "The Journal of Adversarial Examples for Neural Networks", you can probably skip over a lot of the general intro material and assume that someone reading this journal should be familiar with the common challenges you are working on here. For that type of journal, your intro should focus more on the specific limitations of prior research that your work seeks to overcome. Generally these introductions actually end up longer and/or also include a "related works" section to allow you to key in on where your work really stands out.

On the other hand, if you are submitting to a more general interest journal like "Science" or "Nature", you need to do a lot more work covering the basics and usually in a very short space. For these types of journals, the "guff" may actually be valuable information as the readers may not even be aware your field existed before reading your paper. The challenge is getting everyone up to speed in the 200 words or so they give you to talk about it.

Most journals you will be submitting to will be somewhere in between. The general rule of thumb I use is to include in your intro everything your reader would need to know that they wouldn't also need to know for every single other paper in the journal. Even if 10% of the papers submitted to that journal are in your field, that means that 90% of the readers may not be working in that area, so you should at least give enough of an intro to your field to remind them of the things they need to know to understand your paper. That doesn't mean rambling for paragraphs, just enough to set up the more nuanced points so they don't miss the importance of your work.

| improve this answer | |
  • "include in your intro everything your reader would need to know that they wouldn't also need to know for every single other paper in the journal." I'm skeptical. So much publishing is done in broad journals or "mega-Journals." These days journals with narrow scopes have little prestige. I think you can assume readers are a bit more specialized than your answer suggests. – Anonymous Physicist Aug 4 '19 at 9:35
  • 3
    Most machine learning research is published primarily at conferences, not journals. In particular, relatively little ML research is published in The Tabloids (Science, Nature, and their ilk), and most of the CS research that does appear in The Tabloids is actually bad. – JeffE Aug 4 '19 at 16:43
  • @JeffE Can you explain the relevance of you comment? Replace "journal" with "conference" and the same rules apply. – Barker Aug 5 '19 at 13:32
  • 1
    Answers should address the question as asked. Also, there's no such thing as a "more general conference like 'Science' and 'Nature'" (or, for that matter, a "Conference on Adversarial Examples for Neural Networks"). – JeffE Aug 5 '19 at 13:48
  • @JeffE The question does specifically mention journals and not just conferences; I can explicitly mention conferences too if you feel the answer needs it. As for the breadth of conferences, while there are no conference as broad as Science and Nature, conferences do range from broad to more focused so that rule still applies. Also, most ML researchers also present at conferences for the domain they study, so many of the "rules" for a Nature paper would still apply if they were presenting at a medical information security conference, for example. – Barker Aug 5 '19 at 16:22
53

You need to include the introduction to your paper because it might be the first paper in that field the reader is about to read.

| improve this answer | |
16

Your work should be standalone and not rely on the assumption that the reader is familiar with the problem setup and motivation. If you’d like you can slightly shorten the spiel and reference an overview article/several notable papers that cover this. The setup is important because 1. Not everyone reading your paper is guaranteed to know the field 2. Even if they are this gives them some framework as to what they’ll be reading about today.

| improve this answer | |
  • 5
    Reasons why paper should be a standalone: if you are looking for a job, a recruiter might not be from exact same field as you. If you want to hire a research assistant, you want to give them one paper to read (not a list of 20). Besides, even to a reader in your field, standard intro tells them that the paper is part of the same literature. – Bald Bear Aug 2 '19 at 15:34
4

I somehow disagree with the other answers: I am of the same belief that too much of papers or memoirs are inflated with useless information. They are not curriculum books.

The introduction of my PhD thesis was just a few lines more or less stating that if you need an introduction to the problem then you should probably not read the thesis, or have a look at book A and book B before reading.

This was absolutely not intended to show what a genius I am and how uninformed you, the lowly reader, are, but rather to filter out people who will lose their time reading my thesis and expecting some kind of introduction to thermodynamics. I made sure to explain this in these few lines.

This was a thesis and not a paper, but my published papers where similar: the introduction was even shorter.

| improve this answer | |
  • 2
    Fair point in general, but not so much for this specific question. Nobody has yet published the definitive Book A and Book B on adversarial machine learning. – JeffE Aug 4 '19 at 16:46
  • @JeffE: that's an interesting point (and counter argument). What is the introduction for, then? It looks by the wording of the question that there are numerous articles about the subject, so what NN or ML is used for is well-known by the readers. I honestly do not know here. – WoJ Aug 4 '19 at 16:59
  • So this is the kind of thinking I was originally leaning towards (although I’ve been swayed somewhat by other answers). If someone is actually going to try to understand the concepts I’ll present then there’s a lot of background reading they’ll need to do beforehand anyway. For example, I can’t spend several paragraphs explaining the minutia of white box vs black box attacks. The reader will need to go off and understand that thoroughly for themselves anyway. There is not a definitive book for the topic yet, but its uses (safety/security critical) are well documented in vast troves of papers. – dijksterhuis Aug 5 '19 at 12:14
  • There is the situation that happened to me once, where I know of about the general topic of a thesis (say I had 1 class on the subject while at college), but I want to implement/replicate and understand something in the work itself. Hence, I do need a review, and it would be very nice if the thesis would provide it in a consistent manner (i.e. same nomenclatura and notation) with the presented results. Off course, the thesis committee may not need such review, but you have the option to add those introductory notices to an appendix. This doesn't apply to the motivation of the thesis, though. – Mefitico Aug 5 '19 at 16:58
3

Begin every paper with a few words that the reader will immediately accept as uncontroversial to tell the reader what you are on about: 'The Situation'. eg 'Newton showed that gravitation was a universal force governed by an inverse square law'. Very soon after that, preferably the next sentence, state something that complicates the Situation: eg 'Recent experimental results cast doubt on Newton's theory'. You have now raised in the mind of the reader a Question: state it eg 'Was Newton wrong?'. Immediately state your Answer it. eg 'In this paper we show blah blah blah'.

If the reader has got that far, and it is only a few sentences, they will either want to know why you think that your Answer is correct or they will have realised that they are not interested in your paper at all.

That structure: Situation, Complication, Question, Answer enables you to tell your story in a way that will engage the reader throughout.

If it is not already clear from what I have said above, my answer to your question is that you do have to set the scene - that is what the Situation and Complication do, but you write them in a way that sets the scene for your work that you go on to describe, namely your Question, your Answer to that Questions and your reasons for giving that Answer.

If you plunge in then why would anyone want to read your paper? You have to tell them quickly whether the paper is for them. We all have huge numbers of papers that might possibly be relevant to our work. The reader needs to know really quickly whether yours is worth the trouble of reading.

If you say, 'Ah, but that is what the abstract does', then by all means write an abstract on the lines I suggest. If you do that you will question whether you need a section of your paper entitled "Introduction", you could move into more details of the Question (such as previous work on it) and your Answer.

| improve this answer | |
  • 3
    This sounds like advice for an abstract rather than an introduction. – Barker Aug 2 '19 at 23:52
  • Actually it is advice for the construction of an entire paper. see Minto, Barbara, "The Minto Pyramid Principle", 2003" – JeremyC Aug 3 '19 at 7:31
  • 1
    Actually, this is not an answer to the question. – Ratbert Aug 3 '19 at 8:48
  • @Ratbert I thought it was obvious what my answer was. I was wrong; it was not obvious, it seems. So I have added to the answer. – JeremyC Aug 3 '19 at 12:35
  • 1
    @JeremyC The question is not about how to write an introduction in general, but how can one skip the usual stuff and what are the consequences of doing so. So even though your answer may be of interest for some readers, it does not address the initial question. – Ratbert Aug 4 '19 at 7:08
2

I think it's important to have some precise (but maybe concise) introduction to make sure that you and the reader understand the notions, definitions, problems the same way.

Different people think differently, and an introduction would create a common background knowledge/undestanding for correctly understanding your main claims in the following paper.

(Of course, writing and often reading these sections has seemed terribly boring to me, but after some experience as a reader and some little experience as a writer, I came to the idea I explained in this answer above.)

So, whether an introduction is good is whether it helps resolve some misunderstandings/ambiguities concerning your personal viewpoint in your work. There may be some controversies in your field concerning some basic notions and goals, and in the introduction you claim which side you select in a short way, so that no further discussion is due.

| improve this answer | |

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.