My job requires me to read a lot of research papers and implementing them with the standard ML Stack. Seeing that a lot of you have published good quality papers, I would just like to know how you guys come up with good research ideas, as most of the ideas I generate have already been implemented. Don't do research just to do research, but rather focus on things that would be useful to you if they existed, but which don't exist.
That way, even if you get scooped or find out that someone has already done it, you still end up a step closer to where you want to be. And if it is new, it's less likely to be incremental or irrelevant if you actually had a need that it filled when you came up with it.
What kind of research are you thinking of? ML applied to some domain specific problem or method development new topologies etc.
Coming up with a new idea is a long process. Basically a PhD is a research apprenticeship so that might be the best place to start. The new idea comes from a long time exploring the literature and trying things. It doesn't have to be something ground breaking either.
Reading this subreddit you could be forgiven for thinking that everyone is coming up with amazing new techniques, but it is a small sample. As an aside, I wonder if the topology search resnet densenet fractalnet etc for basic image classification is almost exhausted and new advances will come from more theoretical work Author 1: Author 2: Author 3: Author 4: Author SCIgen currently supports Latin-1 characters, but not the full Unicode character set.
We received many donations to send us to the conference, so that we can give a randomly-generated talk. Gil For some reason, this paper was rejected.
We asked for reviews , and got this response. Thanks to the generous donations of people, we went to WMSCI in Orlando and held our own "technical" session in the same hotel. You can download the resulting movie, titled Near Science , below. Movie length: You can read more about the trip here , and check out some pictures here. We hope to set up a better system sometime in the near future. Running the code.
We've been getting a lot of questions about how to run the code. There are quite a few misleading files in the source -- sorry about that. All you need to do to generate a paper is to run make-latex. You can also use scigen. See scirules.
Our plan was to go there and give a completely randomly-generated talk , delivered entirely with a straight face. However, this is very expensive for grad students such as ourselves.
So, we asked visitors to this site to make small donations toward this dream of ours; the response was overwhelming. Related Work.