0
$\begingroup$

I'm trying to classify some 1-D time series data, so I used a simple 1D CNN and fine-tuned the model via Bayesian Optimization (nothing fancy, just used the Keras tuner). And I got very good results (this is on the test dataset, obviously):

enter image description here

But then I saw a method using Continuous Wavelet Transform, we convert the 1-D time series to 2-D scaleograms (frequency vs time graph) and input it to a 2-D CNN. I tried this, and even used transfer learning and Bayesian Optimization, yet, I constantly overfitted the model, and my validation accuracy stuck at 43% while my validation loss got higher and higher (as far as I know, this is overfitting, where the model doesn't learn anything).

So, my question is this: should I focus on the 1D-CNN, or should I try to improve in the second method?

(I have to add one thing: this is for a research paper; so the only reason I started to work on the second method was the fact that my advisor told me that the second method will add "scientific value" to the paper (which I strongly disagreed) and will make the accepting of the paper more "probable.")

$\endgroup$

1 Answer 1

1
$\begingroup$

Academics is what got our industry, and world, to where it is today. With that said, academics is also one of the most vapid worlds I have ever had the misfortune to be a part of. I can still remember my prof wasting my time for the sake of the "story", as if I'm writing a novel, not a research paper.

One of the issues with academics is there's never a "good enough". I don't know if you're overcomplicating things; if the objective is to always get better test results, you can be working on one problem for the rest of your life. To me, those test results seem fine. If you think they can get better, go for it. If you're prof insists on you focusing on "the story", and pursuing avenues merely for their own sake, buckle up.

As for transforming your data into a 2D space via some form of frequency analysis, my eyes immediately lit up. I'm familiar with things in that area, but you mentioned some cool new big words I've never heard of, and got excited to google. I'm sure that avenue has some merit, but my hunch is it has about as much practical utility as flames on the side of a sports car (considering your current results are so good). It will certainly impress the other academics around the water cooler, though.

Your professors live in a world of endless improvement of abstract problems without practical constraint, so when you're in academics, the answer is no, you can never over-complicate. Your sanity may disagree, however. If you want practical constraints, be careful what you wish for, but for what it's worth I haven't had this particular issue in the ocean of practical constraints which is machine learning for a real product. I saw your note, and It reminded me of me. For what it's worth, I much prefer working on tangible problems with real constraints.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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