I'm a CS master student (will be a PhD in three months or so). Today I was at my superviser's office and he had a friend discussing some ideas for their research. Then they mentioned a paper that I knew, that is highly cited, and I also believe is a very good paper. However they both said that it's a crappy paper and I had no idea why! Then also I mentioned arXiv as a resource for papers and they also laughed, as if it was a place for newbies or 'normal' papers. I was embarrassed actually. I would like please to know what are the best sources/places where I can find the best papers. Stuff that I read and people can't say they are 'crappy' papers.
arXiv is, unfortunately, not widely used in machine learning.
You could peruse recent volumes of the Journal of Machine Learning Research for free. It's one of the better journals in the field and entirely open access.
Top conferences like NIPS, ICML, IJCNN, KDD, ... also have freely accessible proceedings which contain some of the most influential publications.
Don't believe anything you see, read, or hear from others, whether of authority, religious teachers or texts [..] Find out for yourself what is truth, what is real.
Essentially, any journal or other source may contain crappy papers. This is because there's no really good way to stop low-evidence or simply false articles being published. Best thing you can do is to try to reproduce result described in the paper and make conclusions. Try to find the truth yourself.
There are, however, some signs indicating that sources are more likely to provide high-quality papers, and most important of them is peer reviewing. E.g. arXiv is simply a collection of preprints that is not actually reviewed by anyone. Essentially, any person with any qualification can upload their work to arXiv, which leads to an average quality of papers being pretty low. Peer-reviewed journals are normally considered more respectable. But still, for any important paper you rely on, it's definitely worth to try to reproduce the result.
(And, by the way, make sure results of your own study are reproducible as well.)
I am not an expert, but AFAIK (because a lot of Machine Learning has a practical approach - always trying to be first to achieve X on benchmark data set?) all the important work is done in conference presentations - the turn around for journals is too slow. So as Marc mentioned - look at NIPS, ICML, IJCNN, KDD...
I think good machine learning papers should be followed with useful and applicable code implementations. Therefore I would recommend paperwithcode
Papers are ranked with Github stars, for the latest and greatest.