How to Read Deep Learning Paper as a Software Engineer
youtube.com
youtube.com
Especially the comment that it takes a week to a month to deeply understand a research paper.
> If I'm trying to understand it deeply enough to mechanistically understand how the methodology work, this can take a few days to a week.
> If I'm trying to reproduce the paper result to incorporate it within my own research project, I might wrestle with the thing for a full month.
> However, I don't do this process with all papers. That would be wasteful. I select which paper have high relevancy for myself and my lab by reading lots of them superficially, then I pin point a few that have disproportionate benefits.
while the "deeply understand" notion might mean something else outside of the scope of the OP (assuming they are not doing this for further research), it seems that practically you spend a few days (not completely) to understand enough.
personally i struggle to find good quality of technical writing in this field than not, and maybe that says about my competence more than anything else. given the empirical nature of tweaking some abstraction layer to have the paper in most cases, there must be a better way to present results.
The reason it takes so long is that academic culture deliberately encourages creating documents that are extremely difficult for people to learn from.
I am confident that in the vast majority of cases the increased understanding that somebody gets from spending 1-4 weeks of effort on a single paper was not actually worth that effort - the same result could have been had in an afternoon of direct conversation with the author of the paper.
Academia really needs to break away from the culture of obfuscating these discoveries! At the very least, any paper worth its salt should be accompanied by an informal blog post that helps explain the discovery and maybe video or audio of the researcher communicating it as effectively as possible.
EDIT: I just spotted that comment, it's a YouTube comment not a note in the video itself https://www.youtube.com/watch?v=nL7lAo95D-o&lc=UgxHeJrOv1g_t...
> If I'm trying to understand it deeply enough to mechanistically understand how the methodology work, this can take a few days to a week.
> If I'm trying to reproduce the paper result to incorporate it within my own research project, I might wrestle with the thing for a full month.
That comment feels very reasonable to me - for actually implementing a new deep learning strategy from a paper a full month feels fine to me.
(I committed the cardinal Hacker News sin here of jumping on an opportunity to share one of my pet peeves, rather than engaging with the information directly!)
Unless you are one of the few experts in exactly the field the author is working on, you will have to work through these references and make the connections to the work.
After this you will have to create a mapping from the clear but h are es to grasp facts of the publication to your internal representation of the topic.
All of these things take time, sometimes you get to skip the part where you have to read the references by consulting a pud thesis that the paper is based on (4 pages to ca. 120 expansion), even than you have to build your internal understanding.
Talking to the author will indeed speed up this process, but that essentially means having an extremely qualified privat tutor.
Isn't it just because of all the knowledge required to actually understand the paper? typically, you wouldn't expect someone without a strong background in the field to understand novel research in an afternoon. This is usually what's a PhD is for: covering the gap between master and state of the art.
While I agree papers should be clearer, and should be accompanied by auxiliary material making the paper easier to understand and apply, I do not agree academia deliberately encourages papers that are difficult to learn from, or of obfuscation. I have a little experience reviewing and writing papers in computer science, specifically security.
As in any field, yes a lot of mentioned but also unmentioned background knowledge may be assumed; and jargon may be used; and it may be written with imperfect style and structure (researchers aren’t professional writers after all, and papers are often written on a harsh deadline). But I don’t think it’s in anyone’s interest, neither the authors nor the community at large, to make the paper inaccessible as a goal and I’ve never perceived this to be happening [1].
Source: have served on many conference PC’s and have been on the receiving end of many reviews, from which I can tell a paper being hard to follow can be grounds for rejection, and a paper being written with clarity and being easy to follow, learn from, reproduce, and build on, is greatly appreciated.
[1] with one exception, come to think of it: needless formalism and mathematical notation, by one research group I remember. I got the impression the intention was to impress/intimidate the reader/reviewers and this was not appreciated.
That's because they are not meant to be used as a learning resource. They're meant to communicate results to other experts who will understand them much faster than you ever could. If you want to learn a new field, start with textbooks and then look for review papers or things like that. Understanding research papers may be your ultimate goal, but don't expect to get there just by consuming these types of papers.
And yet I'm still deconstructing papers entirely within my domain down to the bullet-pointed outline the author used to write it from. We do not need to write like Dickensian wordsmiths and we are not good at it. I'd consider it gatekeeping if I thought people wrote this way out of malice rather than cupidity.
Re winter - there is a difference this time - a hugely useful product (ChatGPT + gpt4 apis) doing billions of revenue. And it's almost all inference revenue.
You can say the same thing about blockchain
ChatGPT is certainly interesting, and some people are paying for it, but I think a) it's not clear how much of that use is also driven by hype, b) of the non-hype use, it's not clear how much will be sustained over the long term, and c) of that long-term use, it's not clear how much is actually of net benefit to society.
Regarding A, an interesting parallel is blockchain hype. There were a zillion blockchain projects, from startups to enterprise efforts. I can't quickly find a reliable number for how much money product vendors, service vendors, and consulting companies took in, but I wouldn't be surprised if that was in the low billions at peak, to say nothing of all the investor and in-house money spent on staff and whatnot. And as far as I know, no non-cryptocurrency success was ever demonstrated. So we can't just suggest that revenue means something will last.
Regarding B, it's a very volatile landscape both in terms of technology and in terms of market. E.g., for a hot minute, everybody loved lively AI generated images as stock photo replacement, but that fad is already past. People are getting a better sense for machine-generated text, too. I know of one company that fired their contract development shop because they kept getting machine-generated communications from the people they were paying to do actual work. Or we could look at Alexa and her kin as an example of something where despite initial excitement, it turns out people just don't care much. Same for VR; since the 1990s the technology keeps getting better and it keeps being a small niche. And we haven't even gotten to how competition and technological improvement will change (or perhaps eliminate) the margins here.
And regarding C, a number of the uses people are paying for are things that are not making the world better. Academic cheating, for example, may be to the advantage of a student who just wants a credential, but it burns a lot of money and makes the world worse. The same applies for people in companies who want to give the appearance of work. Then we have things like spam and influence operations. Over the long term, parasites tend to get squished, however much they can flourish temporarily when they learn a new trick. So that's another slice of AI revenue that can't be counted on.
And I should add that belief among the technically literate has so far been negatively correlated with there not being a winter. In this talk, iRobot founder and noted robot scientist takes a look at that: https://www.youtube.com/watch?v=pgrzEHJTPPM#t=36m55s
In particular, I set the timestamp to his list of AI hype cycles he had seen, 25 of them. Those all had people who understood the technology and believed in it. Maybe it's different this time, but that's what people say in every hype cycle.
I don't believe any image generation service was earnings billions in revenue. Certainly no VR application did. OpenAI is reported to have done over $3 bill in revenue LTM. That's real money, it's all inference money, none of it is hardware. It's all real usage revenue.
The revenue is different. Money talks. And, even though it's anecdata, I personally know dozens of people who use ChatGPT a lot in their job. I used it a lot and get 1000's of dollars a month of value out of it. I worked in VR at peak hype and everyone who worked in it saw the usage numbers - awful, always awful. Not a single person I knew who had a VR headset used it with any frequency. No one.
Regarding blockchain, I was responding to the point that revenue was proof. A shit-ton of money went into blockchain projects. VR has also had significant sales. Both of these are fine analogies to demonstrate that billions of dollars in revenue does not prove a lasting success.
I'm not arguing that nobody uses ChatGPT; I believe they do. My point instead is that it's not clear how sustained the usage will be. E.g., a counter-anecdote: I know of a team who hired a dev who was a big ChatGPT fan. They eventually fired him, because his LLM-generated communication was poor and confusing, and his code was the kind of bad pretty typical of generated code. I'm sure he felt like he was getting thousands of dollars of value out of it up until the point he was shown the door.
One of the things that makes it especially hard to gauge is the extent to which piles of investor money are driving this. During Bubble 1.0, a lot of infrastructure companies had "real revenue" that turned out to be nth-hand investor money. We know ChatGPT has some real users, but how much of that inference traffic is ineffective or unsustainable startups burning up VC cash? That's a question that I expect will take 18 months or so to answer.
Billions of dollars of real usage revenue is an entirely different thing.
Bubble 1.0 is akin to Nvidia revenue - lots of dollars being spent on training, less on inference. Very unclear how much of it ever sees a return. There is a big difference between Nvidia revenue and ChatGPT revenue as far as the issue you are discussing. Inference revenue = real, training revenue (for infra providers) = very speculative.
Even taking it as real, you're the one making claims about how much of it is "real usage revenue". The burden of proof for that is on you. But as is typical for people riding a hype wave, you have not only repeatedly failed to do so, but you also apparently fail to understand countervailing concerns. That you are, in a fashion also typical for the true believer, trying to insist it's my job to prove the hype wrong, marks you as entirely unserious about honest inquiry here. Given that, I'm done.
It's anecdata, but it seems every time this argument is made, it's from a person arguing by analogy rather than first principles, and the person don't really understand how LLMs work and haven't made any real effort to use them. You might not be in this bucket but making an analogy to the get rich quick but no real underlying use case blockchain community suggests you are.
1. revenue doesn't mean much, afaik they still lose money on every query.
2. I think you are using that as a proxy for utility . Companies buying licenses en-masse hoping it will boost productivity is an unproven thesis. Don't see any profit boost in their qtrly reports from chatgpt licenses that they bought.
I use chatgpt and use it daily for coding. it isn't that its completly useless but uses are limited and certainly not worth hundereds of billions of dollars used to literally boil oceans. If someone took away chatgpt from me it not a big deal for me, to be honest. its dumb as rocks for coding, doesn't do even a tiny bit of high level thinking and almost half of it responses are hallucinating garbage.
You use it daily for work. In engineering. How much time does it save you every day? 15 mins? 1 hr? That's a lot at your hourly rate.