video_id
stringlengths 11
11
| text
stringlengths 361
490
| start_second
int64 0
11.3k
| end_second
int64 18
11.3k
| url
stringlengths 48
52
| title
stringlengths 0
100
| thumbnail
stringlengths 0
52
|
|---|---|---|---|---|---|---|
etsayyDGiO0 | particle movement within that flow now interestingly enough I don't know if many of you know this but the Fortran originally you had to start in the seventh column right so space space space space space space and then you would start your statement and things like line numbers and whatnot could go in the first six but that was a throwback to the old punch card days and | 362 | 387 | https://www.youtube.com/watch?v=etsayyDGiO0&t=362s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | with this was in the mid-90s they came out with Fortran 90 and Fortran 90 no longer needed you to have to worry about indentation which was really great right you could just write your program how you wanted and I thought it was fantastic how it unshackled me from having to worry about indentation now I also find it very ironic that in today's stage guido van rossum who invented | 387 | 411 | https://www.youtube.com/watch?v=etsayyDGiO0&t=387s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | python has reshackle me to a programming language that cares about indentation so one other point about this again just to give you an idea of how fast things change so now we're a decade later we're in the mid 90s and I had an IBM workstation and I had received a $10,000 grant for equipment for my research and I spent that entire $10,000 to get a memory module for this workstation that | 411 | 439 | https://www.youtube.com/watch?v=etsayyDGiO0&t=411s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | was 256 megabytes of RAM which was you know a very large Ram back then but $10,000 for that so things change so I got my first job out of school and so everything up to them was great loving life loving programming loving science loving Chemical Engineering and then I got a job in an industry that uses chemical engineering it was the paper industry and my job was to optimize air | 439 | 470 | https://www.youtube.com/watch?v=etsayyDGiO0&t=439s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | and drying systems for drying cardboard right we used cardboard you know they make it wet it's got to be dried and it was I can't say I really enjoyed it you'd have to go and so you know but you can see this is a person you'd have to go and stand up here and you'd have this long temperature measurement that you'd stick in there and measure the temperature of these dryer cans then you | 470 | 495 | https://www.youtube.com/watch?v=etsayyDGiO0&t=470s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | do the same thing with a humidity and you would do it on every single one of those on the back in the front it was hot it was sweaty it was not what I went to college for nine years to do so I decided to upgrade my career and then I went into consumer products making diapers are nappies as they call in some part of the world and slightly more glamorous I suppose but | 495 | 522 | https://www.youtube.com/watch?v=etsayyDGiO0&t=495s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | still not necessarily giving me the career fulfillment that I had always dreamed of and by the way the technology that goes into making diapers is is actually pretty amazing it's just not glamorous work so now it's the early 2000s and I am absolutely bored out of my mind with my work it's just not exciting I I don't enjoy going to work and my company offered a distance | 522 | 548 | https://www.youtube.com/watch?v=etsayyDGiO0&t=522s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | education catalog and I was looking through there and they had a course on computational intelligence and that was three things it was genetic algorithms that was considered computational intelligence back then fuzzy logic which was super hyped up at about that time that I don't know does anyone use fuzzy logic anymore probably not and neural networks so I took this | 548 | 572 | https://www.youtube.com/watch?v=etsayyDGiO0&t=548s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | course now this was pre Coursera days they actually physically mailed you DVDs and you would watch the lectures you know on your television you would do your homework you would email the instructor two weeks later he'd email you back your grades so that was the distance education that was back then I really found it interesting this idea so they taught a very early version of | 572 | 598 | https://www.youtube.com/watch?v=etsayyDGiO0&t=572s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | convolutional neural networks they didn't work well they were very narrow you know they're very shallow and we actually didn't program the so we're still on windows so the instructor would he emailed us a precompiled binary we would run it on our Windows machine which was probably a single-core CPU or back then but even if it had more than one cards we're only using a single core | 598 | 623 | https://www.youtube.com/watch?v=etsayyDGiO0&t=598s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | but we would try to train this thing and and but the the problem was these little pop-can images I should have put a picture and I was blown away that a computer could recognize was it a diet coke was it a dr. pepper was it a sprite I was really fascinated by that the problem was is once I tried to apply this technology to anything in the real world like so in in in consumer product | 623 | 649 | https://www.youtube.com/watch?v=etsayyDGiO0&t=623s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | manufacturing they're taking pictures of every that's being made to do some simple measurements and what and I thought well maybe we can use Ural networks for this could never get it to work so I said well neural networks interesting but not useful so so then fast forward another ten years I'm still bored out of my mind at work right but hey families got to eat and I stumble across this challenge | 649 | 681 | https://www.youtube.com/watch?v=etsayyDGiO0&t=649s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | it was the Merck molecular activity challenge on a platform called kaggle it was fairly simple all you had to do is predict the molecular activity of these you know these molecules based on tens of thousands of molecular descriptors and I've got a PhD in chemical engineering I'm really good at data analysis top prize was $25,000 so I thought to myself this is gonna be the | 681 | 706 | https://www.youtube.com/watch?v=etsayyDGiO0&t=681s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | easiest $25,000 I've ever made in my life right I was literally picking out the car I was gonna buy so the person who won that competition does anyone know who that is it's Jeff Finn so his research lab won the competition using deep neural networks and as Maria said I came in though there were two hundred and thirty six teams I came in actually 23rd but it was from the bottom I came a twenty | 706 | 740 | https://www.youtube.com/watch?v=etsayyDGiO0&t=706s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | third from the bottom and it wasn't because I wasn't trying I was literally trying my hardest and it was it was actually a wait but I mean it was actually it was very demoralizing but it was also a wake-up call because I realized there was all of these things that I had no idea that I didn't know so it was just was like window opened into a whole new set of tools and | 740 | 762 | https://www.youtube.com/watch?v=etsayyDGiO0&t=740s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | technologies that I could learn fortunately another thing happened in 2012 so you probably know Andrew Eng is speaking here in 2012 Andrew hang it was a co-founder of Coursera and I found his course on machine learning and took that course and again I was blown away that here we have a course that would literally have costed thousands of dollars for me to take it's online it's | 762 | 791 | https://www.youtube.com/watch?v=etsayyDGiO0&t=762s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | for free so things were good for a while things were good because you could take that course there of course you could read the few papers that came out Jeff Hanna came out with a few papers but suddenly there is a problem this is just a screenshot of Coursera today and there are more data science and machine learning classes that any person here can take even if it's full-time if you | 791 | 816 | https://www.youtube.com/watch?v=etsayyDGiO0&t=791s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | look at the number of machine learning archived papers that are published every year you can't keep up so on the one hand we live in an amazing time where you can learn the things you need to - like in my case change my career or become good at Kegel on the other hand it's it's too much and you can't keep up and it's information overload so my question is how do you know get given | 816 | 847 | https://www.youtube.com/watch?v=etsayyDGiO0&t=816s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | them given the fact that you must everyone here must keep up their data science skills and the machine learning skills we have to keep up on the one hand on the other hand there's too much to learn so what do you do so the question is how do we keep up this is probably the most important point to understand is you don't you can't you can't keep up with everything new and | 847 | 871 | https://www.youtube.com/watch?v=etsayyDGiO0&t=847s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | machine learning and data science there are people who are deep learning experts who can't even keep up in the deep learning field and once you recognize that it kind of then reframes your mind to say well if I'm not going to learn everything what can I learn and how do I learn it so I'm gonna give three guidelines okay on how you can continue learning - like if you're brand new how | 871 | 896 | https://www.youtube.com/watch?v=etsayyDGiO0&t=871s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | you start your learning journey if you're experienced how you continue and the three guidelines are be sensible be smart and be systematic now I was thinking about branding these the three BS s of learning data science but I'm not sure if that's the best branding you can give me feedback if that's catchy enough so be sensible okay all right I crashed and burned on my first Kago competition | 896 | 922 | https://www.youtube.com/watch?v=etsayyDGiO0&t=896s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | and like I said there was an enormous amount of things I had to learn number one my computer was terrible was four years old had eight gigabytes of RAM and I realized if I want to compete so I started upgrading my computer more Ram Moore's better CPU more disk space more even more RAM better seat a new motherboard right so I started that race that many of us do okay that got me kind | 922 | 944 | https://www.youtube.com/watch?v=etsayyDGiO0&t=922s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | of far I was only using MATLAB in some proprietary software me neither of which you can actually use to win a kaggle competition and at least at the time MATLAB wasn't the best tool so I realized I've got to learn Python so that was another big major learning step the second competition I entered I actually did okay I came in 34th out of 950 34 from the top this time and I use | 944 | 973 | https://www.youtube.com/watch?v=etsayyDGiO0&t=944s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | logistic regression not because it was the right tool it was literally the only algorithm I knew in machine learning so I used it right there may have been veterans uh let's see I know who won like there's some people here I know one that I can't remember I know there's some grandmasters here that actually won that competition so then I learned you know you start seeing XG boost XG boost | 973 | 998 | https://www.youtube.com/watch?v=etsayyDGiO0&t=973s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | is winning everything so then you learn next G pose now uh deep learning I was sure that deep learning was a hype just like fuzzy logic and I avoided learning deep learning over time you know a year later I realized okay this is going to take over the world so I started learning it well that required hardware GPU it required you know I was using Windows so I better go to one to Linux | 998 | 1,024 | https://www.youtube.com/watch?v=etsayyDGiO0&t=998s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | every time I made this shift I felt dumb again I felt stupid I felt like I don't know what I'm doing because it was a huge learning curve so when I say be sensible I want you to recognize that if you start feeling good about yourself and how good you are it's time to jump into something else and to kind of stretch yourself and feel a little bit intimidating again | 1,024 | 1,048 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1024s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | and that's perfectly normal and that's gonna continue to happen III would say anyone who wants to be successful in this area has to get used to feeling insecure about what they know and what they don't know happens to me all the time not I'm I don't like when I say something and it's a mistake or I argue a position and it's wrong on the one hand on the other hand I'm actually glad | 1,048 | 1,072 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1048s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | because right it improves me and I get better so be sensible set your expectations to something that's reasonable so the next thing is be smart and what this means is you have to build a learning plan for you if you go to github all these people say hey a step-by-step learning plan that might be okay it's probably not for you right and the reason is we're all in different | 1,072 | 1,100 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1072s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | places if you go through and somebody's learning plan that is a waste of time to advance not advanced enough you will not do well so be smart so here's what I recommend when you set a learning plan excuse me so say a learning plan say you have five hours a week to learn you all right you say all right every week I'm gonna learn and spend five hours this is how I would recommend doing it | 1,100 | 1,125 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1100s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | number one and I was talking to Pavel earlier this is he's very big on this strengthen your fundamentals so that could be the fundamentals of linear algebra the fundamentals of statistics those things don't change for me most of the time I spend in there is learning Python deeper right learning the language I my favorite book to read and reread is called learning Python it's | 1,125 | 1,150 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1125s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | this thick it takes me forever to get through every time they get going new edition I start over again to refresh those fundamentals the other piece of that is I spend a lot of time learning and relearning pandas every time there's a new addition I learn what's new and I also scan the API for things that oh maybe I didn't realize there was something cool there so strengthen the | 1,150 | 1,171 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1150s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | fundamentals of whatever you need to learn the second big chunk is your core work and your core tools so this is whatever you're doing for your project your problem you know the work you do at work for me is a kaggle data scientist a huge amount of that is domain knowledge so somebody comes to us and says we want to host the competition on earthquake signals I've got to spend time learning | 1,171 | 1,198 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1171s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | about that so I can understand I can ask the right questions when Kaggle was acquired by Google all of a sudden we now and we using our local machines that now we're using GCP VMs right so learning kind of that system stuff I hate to bring this up but we constantly also try to learn the new ways that leakage can enter the data because there's so many thousands and thousands | 1,198 | 1,228 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1198s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | of ways insidious ways for leakage to go and and trying to build our systems around how to how to minimize that that will be different from every one of you and then the last is spending time on the cutting edge now this is where most people make the biggest mistake when they try to keep up with the industry they spend way too much time because there's all this cool stuff there's a | 1,228 | 1,256 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1228s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | capsule networks for neural networks or you know the ordinary ordinary differential equation neural network stuff that was from the nerves conference so some really cool things and I am NOT dissuading anyone from keeping up on those but if you spend the majority of your time learning those and and and trying to keep up the speed you will actually be sub optimizing what's | 1,256 | 1,280 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1256s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | critical for you there are many tools that are going to come out that you'll be able to use right out of the box like Bert the NLP tool that Google pre train model that Google released right that's something that is cutting-edge and you should be able to pull it right back into your you know your tool set so that's approximate you know I would say you don't necessarily need to | 1,280 | 1,304 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1280s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | balance that on a week-to-week basis but over the long term try to to balance that one other point when it comes to learning new things I strongly believe that if you're not writing code and trying it out you're not actually learning it so I fall into this trap I watch a lot of YouTube videos Leyland icon comes out I like to watch as many of the Python videos as possible and | 1,304 | 1,331 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1304s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | unfortunately most of the time I'm not not actually implementing anything I watch it I go yeah that's a great oh that's a great a great idea and then I promptly forget it so if you see something that's important make it a project and get it into your your repertoire so the last vs is be systematic and this is the secret to sustainability this is how you can do | 1,331 | 1,357 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1331s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | this in the long term and not two years from now 200 and I haven't kept up so what do I mean by be systematic I am a big fan of simple checklists and some of you may have read a many years ago I wrote a post mislay it was my most popular forum post I fangled or said here's my standard work that I do for every competition and it literally is I update my Khanda environment I you know | 1,357 | 1,387 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1357s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | whatever so I do any environment updates I create a new repo I download the data I you know so it's just a very simple checklist of things I do so I use multiple machines so I make sure everything syncs out all those machines it wasn't anything that was complicated but the reason I use checklists is because it reduces cognitive friction when you're like oh especially with | 1,387 | 1,412 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1387s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | Kegel we're like oh it's really late I really want to start this competition you don't want to have to think about it so having a checklist allows you to go boom boom boom boom boom and do it without having to think pilots use checklists pilots are some of the most skilled practitioners about what they do and the amount of training they have to go through they use checklists because | 1,412 | 1,432 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1412s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | it reduces cognitive friction if they have a headache if they're tired and make sure they do the appropriate things so so how how do you use a checklist or a system when you're learning for me I have a list of things that I do every day to keep my learning up to date the first thing is super simple I have a bookmark bar for five different archives so archives that write the paper | 1,432 | 1,461 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1432s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | repository archiving topics fluid dynamics data and statistics vision machine vision and machine learning and earth sciences I open those up and I every morning literally before I do anything before I check email I check what those are so that's one of the things I do it's a habit it's sustainable the other thing I do is write any time that any of the PI data | 1,461 | 1,486 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1461s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | the PyCon videos come out I make a list of all the ones I want to watch and I don't get to more than ten percent of them but I make the list and then I worked through them systematically till the next ones come out all right so if you want so as I mentioned there are so many courses it's actually intimidating where do you start what's the right course I have become a big fan of CagA | 1,486 | 1,515 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1486s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | learn not because it's put out by kaggle because I really actually agree with the philosophy what kaggle so if you think of a move if the time to make the presentation to record the presentation to transcribe the presentation to host the presentation it's a very very very very long time scale what learned has done is say we're gonna take some very key important topics and we are going to | 1,515 | 1,545 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1515s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | let you do it in four hours and we're gonna focus on those key important things and get you pushing the buttons and learning these things what's nice about this there's no video actually one of the courses had videos but they're gonna get rid of them is they can freshen them isn't like as much as once a week somebody will ask you a question and they'll realize oh we can make this | 1,545 | 1,565 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1545s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | better so they'll refresh the content so so for example if you if you look at this so it's kind of small but here's the current cago learn everything from python machine learning data visualization sequel micro challenges and the new one machine learning explained ability maybe you know those things but maybe you want a refresher it's really really fun and easy to go | 1,565 | 1,591 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1565s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | through so I'd recommend checking out one so here's like this one on embeddings it's it's I went through it before I I had never got there I went Baba before you through it yesterday I thought oh yeah it makes a lot of sense so cago learned a great place if you ever feel lost go back to Cal go learn obviously there's tons of great stuff out there but if you're ever paralyzed | 1,591 | 1,612 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1591s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | about where you should go go to CAG alert this is an example right so they walk through with words they give you some code you run it it kind of lets you know if you were wrong it gives you hints on very the common errors and it's fun it's interactive it's in kaggle kernels so you can access it anywhere and now that actually keeps track of your progress so you know if you stopped | 1,612 | 1,634 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1612s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | one halfway through it'll tell you which modules you've done all right the other thing that this this kind of surprised me if you go to youtube.com forward slash Kaggle there is an amazing amount of content and it's added to weekly everything from coffee chats to interviews with for example practitioners at Google brain all sorts of pretzeled live coding gets recorded | 1,634 | 1,662 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1634s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | there so this is another great example of a place you can start there's probably too much content there for you to keep up on and that's okay right again give yourself the permission to skip things and focus along those lines there's a point I wanted to make every cago competition that I joined my primary goal was to learn something new so embeddings for example I'm not an NLP | 1,662 | 1,690 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1662s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | expert so if I join a cat go I did a playground who's spooky author so it was like three authors given a sentence predict which one so I did that with the sole goal to learn NLP you may under one to learn okay finally I'm gonna learn convolutional neural networks or ell STM's or some sort of you know whatever you need to go so if you do that no matter where you place | 1,690 | 1,714 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1690s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | you will have one you will have had time well-spent so I want to close up a little bit with this quote it is by far my favorite quote it's a little bit intense so I want to explain it says the illiterate of the 21st century will not be those who cannot read and write but those who cannot learn unlearn and relearn a hundred years ago so many people didn't know how to read or write | 1,714 | 1,739 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1714s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | and that prevented them from being in the workplace in today's world not just in data science and machine learning but almost any vocation you have to learn you have to be able to learn to unlearn and to relearn so that to me is the most important reasons for having yourself a smart plan that's sensible that's smart and that's sustainable so that you can learn | 1,739 | 1,766 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1739s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | whatever skills and do that five 10 20 30 years from now to keep up and have whatever professional success that you want and with that thank you and I'll open it up to questions one second Ronnie Heflin let's see who has who has the mic area what I see a second period thank you oh that's a great question I I'm sure there are but I can't recommend any and the reason being is I just that's not | 1,766 | 1,824 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1766s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | something I looked at does if anyone here know of good data science for kids I you know I would say I don't know for sure but Khan Academy kha in the anatomy is one that has a lot of great content but I don't notice they have data science No okay so sorry great question I'm gonna write that down and we'll follow up yes sir [Music] [Music] yeah it's a really question and and I | 1,824 | 1,912 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1824s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | actually don't believe it's the type of data I believe it's the volume of data and what I mean by that is if your company wants to do a very large data project and you have data in multiple legacy databases that sit in multiple organizations that can be extremely complex and hard to get my recommendation is always to start with the smallest viable scope so try to get | 1,912 | 1,938 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1912s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | a little bit of all the data if you can or just get the data that seems the most reasonable and start doing model building on that the biggest mistake companies make and I've seen this many times is they say well we're gonna have a two-year project to get all of the data collected and then they get all the data collected and they don't have the signal that they need or only five | 1,938 | 1,959 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1938s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | percent of that data was important so I would say just like it was mentioned by grandmasters when your model building you go quick you go small you go fast and then go from there I'd say the same same with data so then then you may say there's there's some very expensive data here that it's worth the time cleaning up it's worth the expense but then you'll find a date over here that's not | 1,959 | 1,986 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1959s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | worth the time okay the biggest mistake is companies say well let's clean up all of our data and that's just waste of time thank you question over here trick question take you back to your fluent that after navier-stokes yes and the reason is is because the navier-stokes is a you know a three dimensional coupled nonlinear equation what the neural networks theoretically can approximate any | 1,986 | 2,031 | https://www.youtube.com/watch?v=etsayyDGiO0&t=1986s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | function so neural networks can approximate the function so I just read a paper last week it was sent to me by a professor in Germany and so there's a there's a type of flow where you have a temperature differential and then you get these circulation patterns and if you look at if you ever seen a close-up of the surface of the Sun you see these like pockets that's what causes that and | 2,031 | 2,058 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2031s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | what they did is they did very very high-resolution simulations of this and then they trained neural networks to identify so segments the the boundaries of this now they get the same results of these expensive simulations but it's orders of magnitude faster and there's a paper that came out a couple of years ago where I think Google was a co-author where like a smoke plume simulate that | 2,058 | 2,090 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2058s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | but then use neural networks to approximate that so now you can model those dynamics so you you you will see more and more of that but the answer is absolutely yes it is exciting work I'll say another thing you're seeing more of this in chemical kinetics so chemical kinetics when you have multiple chemicals it's great it's it's stiff not nonlinear differential equations because | 2,090 | 2,110 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2090s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | you have very fast reactions very slow and people are now using neural networks if you have the system of chemicals we model probably react will see more than yes I am NOT a deep learning expert but I crush because our neighbors are noisy and we do do-dah like 1 million picture is just fine but do you never you are dead picture so great great question so I'll answer that generally | 2,110 | 2,172 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2110s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | first so in general you want to pick the tool that's appropriate there's no reason to do a very very fancy 3d instead if you have a small amount of data and you can do logistic regression then it works fine so I'm a big fan of doing the simplest tool now with neural networks we have the advantage of transfer learning where Google has spent a month training a bank of GPUs on these | 2,172 | 2,194 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2172s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | large networks and so you can use those and actually I'm probably not a hundred but maybe a thousand images because you have all of this pre train models so you know something like that I always say try it right if you have a thousand images kind of fine-tune a pre training model on on 500 see how well it does on the other 500 it's going to be very context dependent I think you know | 2,194 | 2,221 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2194s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | there's some image tasks that are extremely complicated that wouldn't work but there's some simple ones yeah that's a great question and you know to me to be very frank I am very bad at predicting what's next I've always seen what's happening and then act upon that you know there's a lot of people that talk about these things and you know I don't pay a lot of | 2,221 | 2,287 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2221s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | attention to them to be fair i I just watch for tools that I can use in my work I will say this you know data science machine learning AI it's it's not going away in ten years it's not going to wait 20 years I think it's a fantastic field at the end so again is the people that are willing to keep up and whatever happens to be important and learn that I think are gonna do very | 2,287 | 2,310 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2287s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | well for themselves but again I just so I would be out if I knew the answer to that I'd be on the main stage I think we have maybe time for one more question oh yes yes then little stars our afternoon break so we have like three after four minutes do you have any other questions [Music] yeah yeah that so if I understand your question is how do we help those who | 2,310 | 2,368 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2310s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | maybe are discouraged along the path yeah that's that's a hard question of you know so I I came to cattle with a PhD in chemical engineering in a good career and I remember the sense of being very intimidated to ask anything on the forums right I felt I wasn't worthy I felt it was a stupid question that everyone knows it but me I think the more that we tell those stories the | 2,368 | 2,392 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2368s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | people that that's natural that's okay everyone goes through that I think the better right that's why I'm happy to tell my story how poorly I did I tell my story how frightening it was to ask the question on untangle but we can also do a good job I think so tango we try to monitor the forums and I can't say that I do this all the time but if I see somebody that you know they're due and | 2,392 | 2,419 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2392s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | they're trying to say we I'm new to this you know I much as possible try to point them to hate just keep at it don't get discouraged so you know every one of you here can do that you don't have to feel like you're a grandmaster to give people encouragement so I really appreciated that question and regardless of where your level is I think you have the opportunity to do that and again that's | 2,419 | 2,444 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2419s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | what makes Cagle a wonderful community I love the cattle forums there's so much useful content people can enjoy each other a month ago is actually when I met my first Aguilar I had never met somebody from Kaggle but I feel like I know them just by the online interaction so it's a wonderful place and we just have to convince people that they can do that at home and themselves thank you | 2,444 | 2,470 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2444s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | for them hey do you want to ask any more questions okay Pavel I can see your hand what is thank you but it will be the last question that's what your question my favorite game so probably he's maybe not as well known but a parent from Germany Mathias Mueller I liked him because I got ended up with him he was always very helpful he was always very friendly he explained | 2,470 | 2,529 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2470s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
etsayyDGiO0 | things that we easier though easy way for me to understand but but you know it to be fair I there's so many carers the grandmasters that I appreciate for all different reasons like like Abhishek I mean he's just he's just such an interesting character a funny guy as there has a lot to teach people yeah and I can go through the list but I'd say fair and fair and Mattias Mueller from | 2,529 | 2,553 | https://www.youtube.com/watch?v=etsayyDGiO0&t=2529s | Keeping Your Skills Fresh When Everything is Changing | by Walter Reade | Kaggle Days Dubai | Kaggle | |
l0im8AJAMco | good morning everyone the first talk of today is by Jason Lee Corporation University of audacious of departing star customer meetings and over characterization and transition yeah so of course obviously deep learning super successful at blah blah blah it solves a lot of stuff it's successful in a lot of things some people I've even gone as far as claiming it's the new electricity of | 0 | 26 | https://www.youtube.com/watch?v=l0im8AJAMco&t=0s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | course when we have electricity we should look there's a field call electrical engineering the studies it I guess it's kind of a purpose of us being here this summer so deep learning is not quite a science yet it's fairly crazy that things regardless of whether these things where people will try crazy things like these cyclic learning rates learning rate to nice games are pretty | 26 | 50 | https://www.youtube.com/watch?v=l0im8AJAMco&t=26s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | crazy they're fragile even things like changing seeds and it's sort of like we're trying to build a working system by trial and error instead of understanding something like civil engineering and then sort of you know been our deep learner has built this thing but you didn't take a structural engineering class so that's what happened so this talk is trying to about | 50 | 75 | https://www.youtube.com/watch?v=l0im8AJAMco&t=50s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | unkind understand deep learning so motivate of course by this well-known paper I added an optimization at the end I think they forgot that they left out something pretty important there okay so understanding deep learning and we so there's obviously two aspects our goal is to minimize test error we want to do say multi-class classification as usual you can think of as two components an | 75 | 101 | https://www.youtube.com/watch?v=l0im8AJAMco&t=75s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | optimization error and the generalization error something controlled by the algorithm and something simply statistical so there's of course this problem is challenging in both aspects the optimizations non-convex non-smooth the global landscape right nothing is very good about it lots of local minima and just in general it's difficult the reason about the only thing we can see are | 101 | 127 | https://www.youtube.com/watch?v=l0im8AJAMco&t=101s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | really gradients at Hessians and so forth local information and you want to draw global conclusions on the statistical side similarly bad there's more parameters than samples that generalization error literally depends on everything your algorithm you're learning race schedule your architecture initialization scheme algorithm parameters such as momentum so forth okay so how to deal with | 127 | 154 | https://www.youtube.com/watch?v=l0im8AJAMco&t=127s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | optimization luckily this is probably where it's most well understood is simply people have realized that you can always change the model that if one model is difficult to fit we don't have ever believed any models correct or anything like that you just fit a bigger model and bigger models are easier for SPD to fit which is somewhat of an empirical fact I know so you can have | 154 | 178 | https://www.youtube.com/watch?v=l0im8AJAMco&t=154s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | this model with many bad local minima do something and then SGD works so over privatization okay here's an experiment showing this this experiment is essentially replicated from this paper of living Isha Mirror and she'll have shorts so on the left hand side the data comes from a teacher network that's fairly small compact as 50 neurons on the learner side the network has doubled | 178 | 209 | https://www.youtube.com/watch?v=l0im8AJAMco&t=178s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | the number new ons there's a hundred neurons but still the data comes from a network with only 50 neurons so if you try to run SGD on the well-specified architecture which is very sensible assuming you saw through a statistics class if the model comes from the family you want to be well specified Emily's asymptotically efficient da da da whatever by the end of the day actually | 209 | 234 | https://www.youtube.com/watch?v=l0im8AJAMco&t=209s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | fails so statistical efficiency doesn't matter right if you can't get training our low why do you care about statistical efficiency because you cannot find the mo e you run SGD five times it gets five different answers none of which are the global minima I could have done this five thousand times I don't think you would have found the global minimum okay now in stark | 234 | 255 | https://www.youtube.com/watch?v=l0im8AJAMco&t=234s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | contrast simply the same data comes from the same source and I simply change slightly change the architecture the only thing is doubling the number of neurons and I run STD again five times and every single time it's lost value zero and this is also tester this is STD on fresh cause of data this is not training here so SGD is finding the global minima of the population loss | 255 | 283 | https://www.youtube.com/watch?v=l0im8AJAMco&t=255s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | empirically so seemingly over privatization does in fact make optimization easier and in this case even made because it's tests are you could have even extracted some statistical statement here okay so recently in many a series of papers by myself yen's it also briefly talked about this yesterday Trenton has also worked on this I'm summarizing all of these results because they're all kind of | 283 | 313 | https://www.youtube.com/watch?v=l0im8AJAMco&t=283s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | similar enough I don't want to spend time distinguishing they basically say if you have very wide networks and you initialize randomly with appropriate variances that are not crazy gradient descent converges to an epsilon global minimizer as long as the loss is convex your learning rate is small roughly speaking or works for any model so when you're sufficiently over privatized you | 313 | 339 | https://www.youtube.com/watch?v=l0im8AJAMco&t=313s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | get an epsilon approximate global minimizer of the training loss let me sketch a proof of why this could be true of course when you're over parametrized looking at the parameters is not a great idea they're not identifiable there's permutations and all sorts of other invariances that come out that the parameters don't really respect it's a better idea to look at how the | 339 | 362 | https://www.youtube.com/watch?v=l0im8AJAMco&t=339s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | prediction changes so if you compute the change in the loss you'll find that it's related to the outer product of a certain Jacobian matrix or though or the gram matrix as long as this gram matrix is strictly positive definite you get a contraction so then if I could ensure that this thing stayed positive definite then you convert to a global minimizer so why does this the first thing is to | 362 | 390 | https://www.youtube.com/watch?v=l0im8AJAMco&t=362s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | establish that it started positive definite that's pretty clear random matrices are positive definite so this is some sort of just a concentration and perturbation analysis to show that initialization the gram matrix state is positive-definite and then simply when you're very over Prime Christ M is the width of the networks then it stays positive definite there's | 390 | 411 | https://www.youtube.com/watch?v=l0im8AJAMco&t=390s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | always hiding a bunch of stuff it's whatever it's something and then you need am big enough to make this smaller than lambda zero okay so over parameterization essentially what it does is if it's forcing in this initialization scheme it's forcing the grab matrix to stay very stable close to this initialization and thus stay well conditioned and finally you get the | 411 | 436 | https://www.youtube.com/watch?v=l0im8AJAMco&t=411s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | great from this you can extract statements of this type foursquare loss you can say exactly converges to global minima at a log 1 over epsilon rate if this will say logistic loss you could say that it converges an epsilon global minimizer where Absalon is determined by the width if you want in smaller you would need wider because logistic loss is not global we strongly come back to | 436 | 459 | https://www.youtube.com/watch?v=l0im8AJAMco&t=436s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | so many locally well regardless for wide enough networks as long as your loss is any convex function your you can get a statement like this ok so after we so here's a simpler way to interpret these statements if you don't want to read 15 pages of calculating algebra one way to interpret these statements is a statement about the local geometry of the loss function you take your random | 459 | 487 | https://www.youtube.com/watch?v=l0im8AJAMco&t=459s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | initialization draw a ball around it of some norms see for some appropriate seat that depends on the width you set every critical point in B is a global minimizer and there's at least one global minimizer here and Gd store stays in this set so it's a statement about local geometry is saying that locally things essentially look convex as long as you find a critical point in this | 487 | 515 | https://www.youtube.com/watch?v=l0im8AJAMco&t=487s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | local set it will be a global minimizer and in fact you do stay in this set so roughly the global geometry of deep networks is it's probably pretty bad I think there's definitely it should have exponentially many bad local minima local minima with even high values of crane loss however locally there is a global minima locally there exists global minima and you can find a region | 515 | 543 | https://www.youtube.com/watch?v=l0im8AJAMco&t=515s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | in which there are global minima but no local minima no local minima of higher loss yeah one way is you take in under parametrize network and start introducing neurons and then set them to be like to sort of cancel each other configure the ways you add them in a way that cannot help but it's still a local minima I can make at least I can make it a higher order saddle up certain but I'm | 543 | 580 | https://www.youtube.com/watch?v=l0im8AJAMco&t=543s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | not saying you'll find them yeah I'm saying there are ones and certainly not near random initializing oh I think I can always do this I take your architecture I just make a tiny one if you believe there's one in a very small architecture I start augmenting neurons in some useless way so it will look like a standard architecture but maybe the parameters will be in a weird | 580 | 606 | https://www.youtube.com/watch?v=l0im8AJAMco&t=580s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | configuration at least I can make a satellite high order satellite okay so global geometry probably not very good local geometry very good there is a global minima and gradient descent converges to it so we're kind of happy and then oh you think about it a little bit more and then some paper start coming out and then you realize of course locally it's very good locally | 606 | 638 | https://www.youtube.com/watch?v=l0im8AJAMco&t=606s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization | |
l0im8AJAMco | you can write down any function so f theta is my neural network function it's essentially equal to some F 0 it's not important think of f 0 as 0 or very close to 0 some order one quantity a gradient term and a second order term so of course since I'm local the second order term should be thought of as smaller than the delayed the linear term so we kind of just throw it away and | 638 | 667 | https://www.youtube.com/watch?v=l0im8AJAMco&t=638s | On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.