2

So over the last few months, I've been working on an AI model that could classify a skin lesion as cancerous or not using only an image of the lesion. At first, I thought there's no way that I could achieve anything more than 70% accuracy but after tweaking with my program for a long while, I was able to reach an accuracy of over 95% which is definitely an improvement compared to my first tries.

The other day, I was talking to a professor at my local university whom I've written a paper with and showed him my classifier. He told me that I could write a paper on it which had never come to my mind because I believe an accuracy of 95% is not nearly good enough and the methods that I used weren't new, cutting edge methods that could have a huge impact in the AI industry.

Fast forward a couple days, I found all the papers related to classifying skin lesions using ML and DL and I actually found quite a few but they all are vastly different from my model(different hyperparameters, the different architectures, different models, etc) but the accuracy in all of them are pretty close to mine(they were all very close to 95%).

My question is, do you think it's worth it to even try and write a paper about my work or is it pointless and won't help me in the future?

P.S: I've already read a similar question but that was about research in math which is VERY different. In math, the solution is sometimes more important than the finding but I'm not sure if that's the case with AI.

By the way, I'm 14, so I'm not sure whether I can do this independently or not.

Thank you so, so much!

  • What do you mean by “accuracy”? Sensitivity, Specificity, positive or negative predictive value? Or something else? What you have is a diagnostic test, but a diagnostic test can’t be adequately quantified with just one number. – rhialto Nov 1 at 7:30
  • Just a small remark: You say you tweaked your model until you got this number. Are you aware of the dangers of overfitting? Did you make sure to have proper tests in place for that? – Dirk Nov 1 at 7:50
  • 1
    A general AI remark: having three different distinct models that each can classify the same data set to a $95\%$ degree accuracy is inherently more valuable that having just one. It is a common strategy for real life applications to combine multiple classifiers and this often reaches a better classification rate than any of the individual models can manage. – quarague Nov 1 at 10:42
4

As you say, applying well-studied tool X to well-studied problem Y and getting the expected result won't lead to a groundbreaking paper. But not every paper is revolutionary, and many such papers are published, particularly in AI. It is true that professors and other senior researchers try to avoid publishing such papers simply because they have limited time and need to invest it on higher-impact ideas. Only mid- and high-tier papers really count when calculating, for example, tenure decisions.

But you are not a professor, you are 14 and this is presumably your first long-term project. Writing something up and getting it "out there" is highly worthwhile, both so you can practice writing publications and so that you have something citeable when you apply to college and beyond. Even a low-tier publication, or putting it on the arXiv, is pretty darn impressive at your career stage.

I assume this is why your professor recommended that you write up something to publish. Either that, or he thinks the work is more significant than you realize. Since he is familiar with your work and we are not, I would defer to his advice in any case.

0

I would say the general answer to your question is similar in almost every research field. The method to obtain a result is very important and does justify a publication if it is new. One could argue that a new combination of known methods makes a new method.

As you stated, your methods were not new, but you applied known methods to a problem they have not been used for before. That does not exclude publication either. In my field, there are quite a few applied research journals that publish mainly this kind of research. Maybe not highest impact, but fully respected. Is there something like that is AI? Then you could go for that.

And let's be honest: A big fraction of publications is a new combination of known methods to a problem that was addressed in another way before. It is a bit like making music: You only have a few notes that you can combine to a new tune, but if is well done you will not complain about the fact that "already Mozart and Bach used those notes. Okay, they produced a different result, but it is time for an innovation."

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.