@jeremy Can you say more about testing each line of code? e.g. Pitfalls and examples or themes of tests
What’s wrong with the real-time chat here on the forums, @Even? Is there something that you’re finding is missing? Happy to invest in Slack if there’s a clear need.
I find the forums to be much less real time and also much more broken up. A forum is a great place to pose and answer questions and I think it’s an excellent part of the community, but it’s not well suited to quick freeform back and forths or answering of questions in real time.
I jumped on board the MLPeeps slack created by @brendan and had an awesome multiday conversation with @xinxin.li.seattle that’s been a great experience. It could possibly have taken place on the forums, but I feel like with the delays and the extra spacing it would have taken a lot longer to convey the information.
I just want to make sure I understand the issue and confirm that we really need a separate tool, rather than reconfigure this one - so I hope you don’t mind if I ask a few follow up questions?..
Have you tried using private chat groups on the forum here? They are real-time. And we could create a forum on even just a thread for random chat, which also would provide a real-time conversation venue.
Having used Slack a lot and Discourse a reasonable amount I’ve personally found Discourse (i.e. these forums) a superset of Slack’s functionality. I’ve seen conversations between the Discourse devs and various folks on twitter where they’ve tried to resolve similar questions, but it never seems to get far. I’m not sure why, but it seems like no-one has quite managed to clearly articulate either why Discourse is, or isn’t, an adequate real-time chat platform.
What are the delays and spacing that you are referring to?
I want to create the most vibrant possible DL community. If we have it all in one place, but that place doesn’t provide what people need, then it won’t be as good as it can be. But if it’s split over multiple places, then by definition each place will have a weaker community and less momentum, and will also mean more places for people to follow.
We don’t have any funds to pay anyone, but we’re very interested in collaborating with all students to help with (and get help for) their and our projects!
I’m certainly considering it. Perhaps this could be build into the podcast schedule we’re planning - for instance sometimes it could be an interview with someone who has just published an interesting paper, and I could spend time in the podcast working through it myself (just like I have been in class, but with the addition of an interview too). Would that be of interest?
I think PhD requirements won’t be as strict as they seem (except perhaps at large companies, where there’s less flexibility). They just don’t want to waste their time with people who don’t know their stuff, and until now that often meant having a PhD.
Personally I find it easiest to run a different experiment on each GPU. Having said that, Pytorch makes it trivially easy to run multi-GPU training (see the DCGAN lesson for an example).
We certainly would like to, although we don’t have any specific plans or schedules for this. I’m hoping that students will start building on the tools we’ve created themselves as well. I’m already seeing some of the tools discussing and used on the Kaggle forums.
If you or anyone is interested in developing on top of any of the code or ideas from the course, we’d love to help any way we can.
I’m not sure that there are any yet. @timanglade has been working on this - maybe he has some suggestions?
Not really, I’m afraid. It’s one of the most challenging issues in ML. I always try to come up with a strong baseline, by creating a very simple but reasonably effective initial model. Often that means using nearest neighbors, and/or a random forest.
I agree. I hope to develop some good techniques in the coming months. If I’m successful, I’ll certainly want to teach them!
That would be amazing. I think the podcast alone is probably a great start and covers a lot of what I’m looking for. I love podcasts and will definitely be subscribing as soon as it’s available.
Something that might be worth rolling into the podcast as well is @samwit’s suggestion of an overview of new papers and findings in deep learning in every episode. Not only would that round out the episode with additional content, but I’m guessing it’ll drive a lot of interest in the podcast itself.
I know a lot of podcasts that are recorded live as video as well, with a chat audience, so the two could be one and the same.
Right - that’s pretty much what I was thinking.
I haven’t heard one with a chat audience before - could you suggest a couple I could listen to?
The examples I have are a little off topic; I tried to find one in the data science or ML area but there doesn’t seem to be any. The only examples I know of are from gaming, which used to be a hobby of mine before I became addicted to deep learning ;-).
But I would imagine it would work almost exactly as the class does, with a youtube live announcement, a thread on the forum for people to discuss, and then the audio gets put on itunes and the videos can go on your youtube channel.
I did not intend for a paid internship. I actually meant to ask if I can collaborate with any project you are working on. Staying in India , I never had an opportunity to learn from persons like you . I have dedicated my 6 months to 1 year of full time to learn applied deep learning. Thanks for this initiative .
Yes, of course we can collaborate! Perhaps on lung cancer diagnosis?
Sure . I am in. Let me know how can we start.