Comment by nsagent | Hacker News Reader

Comment by nsagent | Hacker News Reader

Parent Full thread

I'm in a similar boat. Not a postdoc yet, but will be starting a postdoc in July (assuming I graduate in May, which I'm on track for).

I've always been interested in the intersection of AI and interactive storytelling. Worked in the game industry for a while, then came back for a PhD when ML really started taking off.

With the current frenzied climate in NLP research, I just feel demotivated. Mainly because I think my research outlook is very different from the mainstream, so I feel my work gets undervalued or ignored entirely.

I spent over a year and a half on my last published research project [1], and it's been largely ignored. Despite having strong reviews after rebuttal, the paper was relegated to the Findings of EMNLP, likely because my research was video game related.

I'm usually the kind of person that focuses on things because I care about them, rather than because others do, but the reality is that hiring decisions in academia (or even industry) require that others value your work. If I truly thought I could do research I cared about and get paid a living, without having to worry about whether others accepted it, I think I'd be much more motivated.

[1]: https://pl.aiwright.dev

My site is a static site hosted on sourcehut, which is having an outage. If it's still down, try https://web.archive.org/web/20240110040908/https://pl.aiwrig...

> Mainly because I think my research outlook is very different from the mainstream, so I feel my work gets undervalued or ignored entirely.

You're not alone.

One of my first papers was in normalizing flows and I focused on a niche area in there (if I said, I'd dox myself. Even this limits the search a lot). Reviewers came back and asked why my images weren't as good as SOTA GANs and just rejected. Half my experiments were on density estimation... The other reviews said I should be applying my methods to GANs instead, but that wasn't even possible because I was explicitly exploiting the distributional properties at each flow step.

My most cited paper is unpublished and the reviews I got back were about why someone would want to train from scratch instead of tuning a large model. Why we cared about such things as small number of parameters or how to quickly train models without overfitting because "bigger models generalize better."

Fwiw, I'd have no issue accepting your work. It looks useful, it advances domain knowledge, and is clearly written. I think a lot of people lose sight that experiments are proxies and that the tools we are working with are more general than the specific applications we demonstrate.

I started my postdoc in August, but the burnout started earlier, a few months before writing my dissertation. Congrats on finishing your degree (almost)!

I also have kind of esoteric interests in my field but it mostly hasn’t bothered me. It certainly doesn’t help with matters though when sometimes I’m trying to motivate myself and go “what’s the point?”

I will say that one thing I’ve learned as a researcher is that it’s hard to know what people will and will not like. I’ve received compliments on some of my least favorite papers I’ve written. My advisor always told this story about how, as a student, he won a best paper award for a paper that almost decided not to publish because they thought the results weren’t strong enough.

Reply on news.ycombinator.com