I’ve recently began working on machine learning, and I was a mathematician before then. Despite such background, as ML is a fast moving area, since I have studied the recent progress considerably and numerous papers, I’m confident in my knowledge. However, I’m neither good at nor fast at implementing my ideas to algorithms. I have no problem with writing pseudocodes, adding details and designing experiments, but it takes bit more time to implement algorithms in Tensorflow and PyTorch for my designed neural network not too slowly. I’m not used to debugging such codes either, since it’s much harder than simple, fast algorithms I had to learn before. Therefore, it takes an incredibly long time to conduct an experiment by myself.
An interesting paper was published 3 days ago, and the author does neither understand its full implication nor have an access to enough computation resources. I can attempt a very interesting experiment, and if successful (which is likely), it would be a great deal for both the author and me. However, due to my incredible slowness, I probably have no chance for it, since others will probably publish before I will. I need someone with machine learning knowledge who is not too bad at coding.
I have three choices: grad students in my university, the aforementioned author and well-known experts in the field. I’ve never met the people in last two options, but both are familiar with the topic. The author doesn’t possess a good coding skill either and failed at coding one thing I’m trying to do, since he is a mere student.
I’m not sure whether I can trust either of them. If I will propose my ideas and coauthoring a paper, since I lack coding experience and it bothers them, there’s no guarantee that, even if they don’t even know the paper (for experts’ case) or my idea yet, they can just claim to have the idea already or just ignore my email, and they can publish the paper based on my idea before I do. Grad students probably can help me, but I’m not sure whether they will show an interest in or familiarity with the topic.
I believe this situation can be generalized. In such cases, what is the best choice? Any advice? By the way, as far as I know, it’s uncommon to just upload to Arxiv a non-theoretical paper with proposal of experiments and generalization of an existing algorithm. Even if I do, it will not give me much credits, since if it does, then proposal papers would have abounded.