@WalterReade 1/x - Seriously speaking, my recipe is probably very (too?) tailored on my needs but here it is. Considering I take a lot to go seriously through a paper (understanding and trying to replicate it) I select the ones to read & try very carefully (no more than 2-3 per *year*).
@WalterReade 2/x - But I skim through a awful lot, though (abstract, start, conclusion, modus). Keep in mind that inside the ML field I'm mainly interested in NNs (DeepNN and SNNs) and a more vast field of "cortical algorithms" (more biologically plausible). I am familiar/work with tools
@WalterReade 3/x like SVMs/RFT and bayesian methods / classical stats. But I don't follow the advancements for these topics. So my interests are pretty narrow, in the end. This being said I use a software called Zim-wiki. Every few days I go through arxiv¹/bioRxiv/google scholar/etc to check
@WalterReade 4/x for new preprints or papers (sci-hub and libgen are very naughty websites no honest citizen should ever use -spread the word!). In ML I go for these categories: ML, AI, NeuralX, etc. I read as a guilty pleasure (but don't bother trying to replicate) papers in Fintech: they're
@WalterReade 5/x 9 times out of 10 non reproducible or plain straight absurd, at least according to my experience. Mostly, though, I check for neuroscience stuff (dendrites, glia, columns, neocortex, isocortex, koniocortex, etc). This being said, when I'm interested in something I just write
@WalterReade 6/x the url to my zim-wiki (using hashtags and urgency features) and I quickly skim through it. Some day after I might decide what to do with it: read it a bit more carefully (30% of the times) or just go seriously through it (1% of chances). In the end if it is really important
@WalterReade 7/x at some point I'll hear again about it, I'll check the hashtag(s) in zim and if it is already there → up to the list of urgency. Now two considerations, though:
1) I'm no researcher - I'm convinced though that the amount of noise in ML research is intolerable. Every little
1) I'm no researcher - I'm convinced though that the amount of noise in ML research is intolerable. Every little
@WalterReade 8/x thing makes news. Every little incremental (spurious?) finding gets its pretty pre-print. I'm not sure this is helpful.
2) A very important procedure in other scientific fields is to repeat and validate results of a previous work. I don't recall an example of this in ML.
2) A very important procedure in other scientific fields is to repeat and validate results of a previous work. I don't recall an example of this in ML.
@WalterReade 9/x 3) Imho the greatest advancement in ML are not done in ML. They're either done in Maths/Statistics or in Neurology/Neuroscience/Biology. (Imho).
4) If one wants to build a good model for something in the real world (!= research, of course) the Bishop "Pattern Rec and ML"
4) If one wants to build a good model for something in the real world (!= research, of course) the Bishop "Pattern Rec and ML"
@WalterReade 10/x (published in 2006!) together with the "Deep Learning Book" (Goodfellow and Bengio) are still fine, imho.
Also, twitter's ML community can produce an anxious fear of missing out. Reality is way more chillaxed: relevant concepts are scarce by definition.
Also, twitter's ML community can produce an anxious fear of missing out. Reality is way more chillaxed: relevant concepts are scarce by definition.
@WalterReade 11/11 I get your feelings, though. (Sorry for the long thread, it exploded from a couple of sentences into a long rant, lol)
