08:04 < ShikharJ> Can I ask something unrelated to mlpack but concerning to ML and AI in general?

13:33 < zoq> ShikharJ: Sure, anything.

14:09 < rrahul> whois zoq

14:10 < rrahul> sorry for that

15:15 < ShikharJ> zoq: If supposing you had to learn ML and DL from scratch (in hopes of pursuing a Ph.D. later, under a top researcher later), how would you go about that? I'm sorry it sounds a bit weird, but it is my query.

16:43 < zoq> ShikharJ: I don't think this is a weird question; I don't think there is a single path you can go to get what you want. For some people it's all about publishing, for others it's applying ML methods, or participating in competitions. I really like the answers from the discussion: https://www.quora.com/What-are-must-dos-for-a-PhD-student-in-Machine-learning

17:26 < ShikharJ> zoq: Thanks for this. If I can ask, what fields of mathematics should one have a grip on (apart from Linear Algebra, Probability and Optimization Theory)? Some suggest multi-variate calculus and/or differential equations, but I haven't yet come across a domain in ML requiring that knowledge.

17:29 < ShikharJ> I certainly do read the research literature and the classic books (Goodfellow, Bishop, Kevin Murphy etc.) every now and then, but is there a specific resource that I should be making use of in addition to these?

18:02 < zoq> Each field/subfield has its own bits that can be learned along the way, so if you like to look into for instance physic models, differential equations are definitely something you could look into.

18:02 < zoq> What I like to do is to, at least glimpse over the papers that are coming out that, at least on the one I think look interesting; To manage the overwhelming amount of papers I usually use: http://www.arxiv-sanity.com/

19:55 < daivik> Hi everyone, I'm looking to contribute to mlpack for GSoC 2018. I have built the project from source and have been looking through the codebase and some tutorials; and trying to write some simple programs by myself over the past few days. I have also seen the GSoC 2018 ideas page and am quite interested in contributing to the Essential Deep Learnin

19:55 < daivik> g modules project on there -- are any of the algorithms metioned there (GAN, spike and slab RBM, Bidirectional RNNs) open (that is, noone is working on them yet so that I can maybe start looking into those)? I also notice that there is already an HMM implementation in the library and was just wondering if it would be helpful to have MEMM and CRF im

19:55 < daivik> plementations in there as well? If so, then I would be very interested in working on those - but I dont know if that is good enough for a complete proposal by itself? Any advice on which direction to get started in would be greatly appreciated. Thanks a lot!

20:21 < zoq> daivik: Hello there, there is an open PR for standard GAN: https://github.com/mlpack/mlpack/pull/1066 and I think ShikharJ is working on the remaining issues here: https://github.com/mlpack/mlpack/pull/1204, if you are interested on the same you could ask Shikhar if you could colloborate on the changes, but don't feel obligated. Besides that there is also an open PR for the RBM:

20:21 < zoq> https://github.com/mlpack/mlpack/pull/1046

20:22 < zoq> daivik: Anyway, there are other interesting GAN, RBM implementations which aren't covered yet, so feel free to work on that or write an application for the same.

20:22 < zoq> daivik: The HMM/CRF idea sounds interesting if I remember right, Ryan (rcurtin here on IRC) already worked with CRF's and can probably give some input on the idea.

20:22 < zoq> daivik: He will response on this one once he has a chance.

20:36 < daivik> Thanks for the quick response. I was also wondering if genetic algorithms could also be worth looking into (since one of the project ideas is PSO) -- atleast for unconstrained problems? Although I think that maybe the chromosome representation step would be a problem -- since it depends on the person who designs the experiment. I'm sorry for just t

20:36 < daivik> hrowing random thoughts out there, please feel free to shoot them down anytime...

20:40 < zoq> Genetic algorithms is definitely an option, I would love to see a really fast NEAT implementation inside mlpack, of course there are other methods which are interesting as well.

20:44 < zoq> Also, I agree, finding a good representation is challenging, especially if you aim for dynamic topologies.

20:46 < daivik> Hasn't someone already done NEAT? -- I think I saw it somewhere..maybe in a PR

20:50 < daivik> https://github.com/mlpack/mlpack/pull/1105 --> this has not been merged yet I think. Are you suggesting a faster implementation?

20:52 < zoq> Yeah, there is an open PR, there is a hidden issue; at some point the structure does not evolve anymore.

20:56 < daivik> Okay, thanks a lot .. I guess I'll just pick one of these then? Any particular one you would suggest for a newcomer to mlpack?

20:59 < haritha1313> zoq: Hi, I found your suggestion of the research paper on 'Deep Tensor Convolution on Multicores' in an issue. I read the paper and found it really interesting. Do you think it will be a good idea to work upon?

21:00 < haritha1313> oops, seems like I popped up in the middle of another conversation :D

21:27 < zoq> daivik: If you like you can take a look at any idea, but don't feel obligated.

21:29 < zoq> haritha1313: Hello, this is a good idea to dive into the codebase for sure, but it's also somewhat complex. So please don't hesitate to ask if you have any question.