Hacker News Reader

Comment by PragmaticPulp | Hacker News Reader

Is your goal to help advance the science and our general knowledge? Publish the code. You don’t even need to clean it up. Just publish. Don’t worry about coding style nitpicks. Having the code and data available actually protects you from claims of fabrication or unseen errors in hidden parts of your research.

On the other hand, if your goal is only to advance your own career and you want to inhibit others from operating in this space any more than necessary to publish (diminish your “competitive advantage”) then I guess you wouldn’t want to publish.

View on HN

Not sure if posting the paper only is even the best move. I personally never work with papers with no code published. Just not worth the effort to reproduce them, when I can use the second SOTA for nearly no performance penalty and much less effort.

All the groundbreaking papers in deep learning in the last decade had code published. So if you're aiming for thousands of citations, you need code.

Can confirm.

I personally look at any paper without code with great suspicion. The reviewers certainly did not try to reproduce your results, and I have no guarantee that a paper without code has enough information for me to reproduce.

I always go for the papers with code provided.

Agreed, reproducibility helps a lot, and it is very easy to get details wrong when reimplementing. Having the source code is a bit plus.

As a reviewer I have reproduced results with my own independent plasma simulation code. And I have had a reviewer write in a report about my paper "result X seemed strange, but I ran it with my code and it does it too. I don't know why, but the results is valid". In my opinion that is even better than just rerunning the same code.

> All the groundbreaking papers in deep learning in the last decade had code published. So if you're aiming for thousands of citations, you need code.

I am in this field and I would say less than 10% of the top papers have code published by the author, and those are most of the time another 0.1% improvement in imagenet. All the libraries that you generally use are likely to be recreated by others in this field. Lot of most interesting work's code never come out like alphazero/muzero, GPT-3 etc.

This is very domain specific. OP said it is not the norm to do publish code in his field. I have a PhD and in my field it is the same. So much so that I can't think of any paper in my field that has code published. Therefore, a paper with no code would not be at a disadvantage.

Personally, it is a pet peeve I have about my field. But there is no incentive for a new researcher to publish code as it decreases barriers to entry. As much as it's nice to say that researching in academia is about progressing science, as a researcher, you are your own startup trying to make it (i.e., get tenure).

Reply on news.ycombinator.com