1. Understand the title 2. Don't pay much attention to authors lol 3. Read abstract well and form hypothesis 4. Look at pictures to understand solution 5. Read introduction carefully 6. Skip related work 7. Read everything else and skip things that don't seem like they are part of the general idea 8. Look at the results and try to prove them wrong or get convinced that solution works 9. Glance over everything and make sure all your questions were answered 10. Read it a couple of times or take breaks to have a good understanding
1. **Read the title** and make an opinion of what's in the paper (e.g., the area, the task) 2. **Read the abstract well** and form a hypothesis of 1. What's new in the paper? 2. Do you have a clear **overview** about what the paper is all about? 3. **Look at the images** and extract a set of "questions" about what is not clear about their method from the images. Now your job is to answer these questions by reading the paper. 4. **Read the introduction carefully** to get a high-level understanding of the paper. 5. **Read the method** aiming to answer your "questions" about the paper. Focus on understanding only the things relevant for the story (i.e., to understand the contribution). 6. **Read the experiments** to convince you that the show results are caused by their claim. Be aware that the experiments highlighted are the best scenarios and are fully hyper-parameter tuned. 7. **Make sure you answered all your questions. ** Did the authors convince you that their story has the effect that they claim?
I think this video does not only give a guide to how we can read a paper fast, but also how to write a good paper to help readers understand it more efficiently.
I think reading papers is like building your cache of knowledge. The first paper I read took me ~40 hours because I had to google each line and then google each line of the explanation. Its like fractal googling. Once, the cache is built it gets faster. Relevant XKCD: xkcd.com/739/ . 22:05 Related work is useful for practitioners who are looking for alternative solutions to a problem. Very often google search or even google scholar search does not return relevant papers. This human curated citations helps quite a bit. Some papers have code, some dont. So, if I see another paper which is solving the same problem but has code on Github, I go and read that paper instead. . I find taking notes and writing a summary of the paper helps in retention. It also helps me quickly search through my archive of summaries to find relevant papers.
Yes I didn't want to say the related work section is irrelevant, but for the purpose of me understanding the paper, it's most often useless. I wouldn't advise leaving it away when you write a paper :)
This was so helpful! I love that you picked a particular paper to read through instead of just telling about an abstract process you follow. That really cemented it for me. Thank you for making this video!
It's weird how we all do this differently. I know that I can get about 85+% of the context in the authors' heads by reading the related work section, where they inevitably state that the paper is different from others because of XX or YY or ZZ. From there, it's usually trivial to figure out what they intended, so even if the notation is awful (scattered all over) and the explanations are threadbare, I can still understand their MO. It's gotten to the point where I read the abstract and then just skip right to the related work section to figure out what they intended to accomplish to set themselves apart. This only fails when the paper is super novel (and bleeding edge is usually fun to read but not super useful) or when there is no related work section. Your ability to put up with the tedium of deciphering their notation is admirable. I don't have time to get in the heads of a few graduate students attempting to communicate a method, so I typically look at the pictures, the algorithms, and then connect the dots from my related work priors. "We changed five letters in the algorithm but it's identical to the previous one except for one little part" is too frequent for me to put up with trying to follow along. Your method is basically "I read the whole code base to understand what's going on" and mine is "I did a diff of this with the prior code". Both work, and mine occasionally explodes spectacularly, but usually it's faster.
Thanks a lot for your extensive explanation of the method you follow. I am a masters student in visual computing and this helps me a lot in changing the approach in which I follow a paper.
Really awesome information! I could probably have benefited a lot from this when I was first learning to read ML papers. I will definitely share this with any of my friends who are beginning their research careers. For anyone starting out, I found it really useful to take a few well-respected and well-written papers in any area and read through them extremely thoroughly making sure I understood all of the details (even if this does take a few days initially). I found that once you understand some of the most popular ideas, it gets a lot easier to speed read new papers, especially those that recycle old techniques in new ways.
Very nice! A general rule for Writing papers (though I'm an oceanographer not an ML researcher) is that the "pictures"( with captions) should tell most of the story in themselves.
Thank you very much, it was really helpful. Could be "obvious" or "simple" once you know it, but for me, this video taught me a lot about the strategy in doing so, where to put more attention (and when), and what should (and not) expect to achieve.
Personally I actually like the Related Work section -- if I don't fully grasp what the paper is about, I can at least learn how it relates to other papers I might know. Also, if the paper is on a topic that I'm less familiar with, this is a great place to get references for further exploration.
I found it very helpful the way it is, thank you very much! About more of this kind of videos: If you find some useful "how to read a paper" tips along the way that you think could be beneficial to others, too, then I will be very happy to watch that video!
I was waiting to know how you read the Broader Impact section... XD. "I go with a totally unbiased opinion about the section and consult a few people from the humanities department to see if the authors have convinced me"..."I am lying I read it to mock it"
The paper used the wrong font, obviously it's trash :-P Jokes aside, this is great! If I'm reading it seriously and not just skimming for something, I will often transcribe the paper I'm reading into bullet-point form (mostly verbatim, not really summarized), which generally helps me attend to it better - especially with all the breaks I take to think about something, its implications, and how it would be coded (etc.)
Found your channel by accident. Watched your serials of video about DETR. It was great. You enjoy a lot during coding. Can you make more video about coding? Hope I can enjoy coding too...
Dang, just about every paper is about Transformers now, isn't it? I wonder, do you think it'd make sense to just train a transformer on any given data set (doesn't really matter that much which one, just, like the idea in general) using that idea from DADS (Dynamics aware discovery of skills)? - Just have the various attention heads be the different tasks according to the DADS algorithm or something like it. It *should* learn random, mutually unrelated, but individually informative tasks which then could be individually investigated, or combined in some way to complete a specific task. Would be really interesting to see what such a situation would do to the language domain, say. What would each head even put out without a given goal?
Could you detail a bit more your usual bias towards thinking experiments are crap? I have somewhat the opposite feeling and am curious what data points you have
I am new here, I got to know about this channel from twitter. Every video looks interesting but i can't understand. Can anybody explain as a beginner( still building my first project) in machine learning how can i get benefit from this channel? And does this channel have any watching index which shows beginner level to advanced level videos? Thanks any help would be great. :)
Always nice to see how others do it, thanks Yannic! Btw. a relevant read from Gwern, titled "Humans Who Are Not Concentrating Are Not General Intelligences": www.gwern.net/docs/www/srconstantin.wordpress.com/486dba34fb7c61678ed10541ef4b71efc0c56918.html I also find myself skimming and thinking I understand the text (not necessarily a paper) but the fact is I don't. Gwern showed nicely how once you get into that mode GPT-2/GPT-3's text seems indistinguishable from the human-written text. Anyways, keep up the good work buddy! Btw. how old are you? I saw you're with ETH Zurich, I've got some friends there and Marc Pollefeys is on my team at Microsoft. I got to chat with him here and there like at ICCV 19' last year as well.
Hello Yannic! Your transformer-series is my favorit! Thanks for your incredible ability to explain complex in easy understandable way due to yourself profound understanding. I wonder did you also read the transformer application on video action recognition? arxiv.org/pdf/2102.00719.pdf If you feel this is interesting, looking forward to your insight in your next tutorial video!
![[Classic] Deep Residual Learning for Image Recognition (Paper Explained)](/img/n.gif)
1. Understand the title
2. Don't pay much attention to authors lol
3. Read abstract well and form hypothesis
4. Look at pictures to understand solution
5. Read introduction carefully
6. Skip related work
7. Read everything else and skip things that don't seem like they are part of the general idea
8. Look at the results and try to prove them wrong or get convinced that solution works
9. Glance over everything and make sure all your questions were answered
10. Read it a couple of times or take breaks to have a good understanding
1. **Read the title** and make an opinion of what's in the paper (e.g., the area, the task)
2. **Read the abstract well** and form a hypothesis of
1. What's new in the paper?
2. Do you have a clear **overview** about what the paper is all about?
3. **Look at the images** and extract a set of "questions" about what is not clear about their method from the images. Now your job is to answer these questions by reading the paper.
4. **Read the introduction carefully** to get a high-level understanding of the paper.
5. **Read the method** aiming to answer your "questions" about the paper. Focus on understanding only the things relevant for the story (i.e., to understand the contribution).
6. **Read the experiments** to convince you that the show results are caused by their claim. Be aware that the experiments highlighted are the best scenarios and are fully hyper-parameter tuned.
7. **Make sure you answered all your questions. ** Did the authors convince you that their story has the effect that they claim?
I think this video does not only give a guide to how we can read a paper fast, but also how to write a good paper to help readers understand it more efficiently.
Andrew Ng has a nice lecture on how to read a paper, which I was reminded of because he also advises looking at pictures first
link please
@@awimagic czcams.com/video/733m6qBH-jI/video.html
may be this one? czcams.com/video/733m6qBH-jI/video.html
This is super-interesting. It's good to see the "model" you have of a paper.
I think reading papers is like building your cache of knowledge. The first paper I read took me ~40 hours because I had to google each line and then google each line of the explanation. Its like fractal googling. Once, the cache is built it gets faster.
Relevant XKCD: xkcd.com/739/
.
22:05 Related work is useful for practitioners who are looking for alternative solutions to a problem. Very often google search or even google scholar search does not return relevant papers. This human curated citations helps quite a bit. Some papers have code, some dont. So, if I see another paper which is solving the same problem but has code on Github, I go and read that paper instead.
.
I find taking notes and writing a summary of the paper helps in retention. It also helps me quickly search through my archive of summaries to find relevant papers.
Yes I didn't want to say the related work section is irrelevant, but for the purpose of me understanding the paper, it's most often useless. I wouldn't advise leaving it away when you write a paper :)
AWESOMEEEEEE!
This was so helpful! I love that you picked a particular paper to read through instead of just telling about an abstract process you follow. That really cemented it for me. Thank you for making this video!
It's weird how we all do this differently. I know that I can get about 85+% of the context in the authors' heads by reading the related work section, where they inevitably state that the paper is different from others because of XX or YY or ZZ. From there, it's usually trivial to figure out what they intended, so even if the notation is awful (scattered all over) and the explanations are threadbare, I can still understand their MO. It's gotten to the point where I read the abstract and then just skip right to the related work section to figure out what they intended to accomplish to set themselves apart. This only fails when the paper is super novel (and bleeding edge is usually fun to read but not super useful) or when there is no related work section.
Your ability to put up with the tedium of deciphering their notation is admirable. I don't have time to get in the heads of a few graduate students attempting to communicate a method, so I typically look at the pictures, the algorithms, and then connect the dots from my related work priors. "We changed five letters in the algorithm but it's identical to the previous one except for one little part" is too frequent for me to put up with trying to follow along. Your method is basically "I read the whole code base to understand what's going on" and mine is "I did a diff of this with the prior code". Both work, and mine occasionally explodes spectacularly, but usually it's faster.
Thanks a lot for your extensive explanation of the method you follow. I am a masters student in visual computing and this helps me a lot in changing the approach in which I follow a paper.
Really awesome information! I could probably have benefited a lot from this when I was first learning to read ML papers. I will definitely share this with any of my friends who are beginning their research careers. For anyone starting out, I found it really useful to take a few well-respected and well-written papers in any area and read through them extremely thoroughly making sure I understood all of the details (even if this does take a few days initially). I found that once you understand some of the most popular ideas, it gets a lot easier to speed read new papers, especially those that recycle old techniques in new ways.
Very nice! A general rule for Writing papers (though I'm an oceanographer not an ML researcher) is that the "pictures"( with captions) should tell most of the story in themselves.
Like this explanation because it is exemplary not just of a good papers-reading technique but a good papers-writing approach as well.
How you respond to our inner desire to learn how to learn, amazing very helpful. Found the opinion/insight on related work very useful
Thank you very much, it was really helpful. Could be "obvious" or "simple" once you know it, but for me, this video taught me a lot about the strategy in doing so, where to put more attention (and when), and what should (and not) expect to achieve.
22:18 "First of all you cite a bunch of your friends...". Lol!!! You reverse engineer and share the secret of the games :D
Personally I actually like the Related Work section -- if I don't fully grasp what the paper is about, I can at least learn how it relates to other papers I might know.
Also, if the paper is on a topic that I'm less familiar with, this is a great place to get references for further exploration.
Also it is a good place to have an understanding of what is going on in the field. Some papers are really detailed
Thank you, I like how you make your videos very informative and fun to watch
Thanks for taking the time to do this.
This is very informative and concrete, which is extremely nice. The long form works really well for you, please continue!
I found it very helpful the way it is, thank you very much!
About more of this kind of videos:
If you find some useful "how to read a paper" tips along the way that you think could be beneficial to others, too, then I will be very happy to watch that video!
Incredibly helpful
I was waiting to know how you read the Broader Impact section... XD. "I go with a totally unbiased opinion about the section and consult a few people from the humanities department to see if the authors have convinced me"..."I am lying I read it to mock it"
Some things remain trade secrets ;)
I watch your video first. :)
writting notes really helps. at the end if you try to read the notes you'll pick all the parts you didnt understand.
Bond.
Much needed! Thank you! 😇
Thank you for this nice explanation.
Very useful, thanks so much for sharing.
Great video. Thank you for doing this.
Very useful, thank you for sharing
This video is a godsend
Thanks alot man 😁
Great video! Learnt a lot
42:28 « And then I gloss over the _abstract_ » Ehm… Appendix! 😜
Well they're almost the same :D
Great video!
This is very nice, thanks!
"first I go to the pictures... " XD
same
The paper used the wrong font, obviously it's trash :-P
Jokes aside, this is great! If I'm reading it seriously and not just skimming for something, I will often transcribe the paper I'm reading into bullet-point form (mostly verbatim, not really summarized), which generally helps me attend to it better - especially with all the breaks I take to think about something, its implications, and how it would be coded (etc.)
Thank you very much, this is very helpful
Pictures < 3
Nice video, thx!)
Found your channel by accident. Watched your serials of video about DETR.
It was great. You enjoy a lot during coding.
Can you make more video about coding?
Hope I can enjoy coding too...
Yaay Yannic....
Bless you
EM LOVIN IT...
Hi
Can you make a video on how you manage time on research projects and then make videos like these also. Community can learn from that.
Thanks!
Yes. I don't manage my time. I miss things constantly. I'm a mess :D
Could you share how you take notes? Which app do you use? Do you read in a laptop, kidle, tablet? Do you use a digital pencil?
I'm on a surface tablet using OneNote
What Software do you use to edit PDFs?
Thanks. Great work. What is the tool you are using for marking and notices in pdf?
OneNote
Thank you! Very helpful tutorial. And I'd like to know what is your app for taking note of papers?
I personally don't take notes when I read.
@@YannicKilcher Alright. But still thanks for your reply! Feel like a fans now
Dang, just about every paper is about Transformers now, isn't it?
I wonder, do you think it'd make sense to just train a transformer on any given data set (doesn't really matter that much which one, just, like the idea in general) using that idea from DADS (Dynamics aware discovery of skills)? - Just have the various attention heads be the different tasks according to the DADS algorithm or something like it.
It *should* learn random, mutually unrelated, but individually informative tasks which then could be individually investigated, or combined in some way to complete a specific task.
Would be really interesting to see what such a situation would do to the language domain, say. What would each head even put out without a given goal?
Yes, that's a nice idea. Goes a bit into the direction of Bengio's ICLR talk
@@YannicKilcher czcams.com/video/GYqSNv_j1-Y/video.html this one? Will check it out, thanks!
29:30 good point ;)
Which app do you use to write on PDFs in such a manner?
OneNote
@@YannicKilcher Are you using a tablet to read papers?
Could you detail a bit more your usual bias towards thinking experiments are crap? I have somewhat the opposite feeling and am curious what data points you have
I and many others I've known have tried too many times to replicate papers at the beginning of our PhDs.
I am new here,
I got to know about this channel from twitter.
Every video looks interesting but i can't understand.
Can anybody explain as a beginner( still building my first project) in machine learning how can i get benefit from this channel?
And does this channel have any watching index which shows beginner level to advanced level videos?
Thanks any help would be great. :)
You should probably go to other channels first, search for introduction to machine learning
Which application do you use for reading papers? I think this was OneNote
I think it is Notability on iPad.
It's OneNote
Always nice to see how others do it, thanks Yannic! Btw. a relevant read from Gwern, titled "Humans Who Are Not Concentrating Are Not General Intelligences": www.gwern.net/docs/www/srconstantin.wordpress.com/486dba34fb7c61678ed10541ef4b71efc0c56918.html
I also find myself skimming and thinking I understand the text (not necessarily a paper) but the fact is I don't. Gwern showed nicely how once you get into that mode GPT-2/GPT-3's text seems indistinguishable from the human-written text. Anyways, keep up the good work buddy!
Btw. how old are you? I saw you're with ETH Zurich, I've got some friends there and Marc Pollefeys is on my team at Microsoft. I got to chat with him here and there like at ICCV 19' last year as well.
Can you implement a paper in a video ?
Hello Yannic! Your transformer-series is my favorit! Thanks for your incredible ability to explain complex in easy understandable way due to yourself profound understanding. I wonder did you also read the transformer application on video action recognition? arxiv.org/pdf/2102.00719.pdf If you feel this is interesting, looking forward to your insight in your next tutorial video!
Haven't read it yet, thanks for the suggestion
LOL, this kinda how I read papers, except for like experience and understanding the math.
This is the opposite of a "Two minutes papers" video lel
read the title, abstract, then pics...
And some call this reviewing... At least this is interesting on how to fool noobs reviewers.
The most part of algorithm you are talking about are stealing from Andrei Munteanu 1989
Maybe you should write a paper on how to read a paper.