1D convolution

[wstang35]

Hi Abhishasike！

I read your post on reddit and your blog about dilated CNN, it is really helpful.

I just started using 1D convolution, and I am working on a EEG project, in which I try to use 1D CNN to extract some waveform features from raw EEG inputs.

I will try dilated convolution later on my network, and see if it works on my problem.

By the way, I see your comment in another post in Reddit, and you metion about some current advances in time-series prediction with CNN's. like auxillary data concerning relationships between significant points of the signal.

So how can I find some papers about metion about some current advances in time-series prediction with CNN's, any papers recommended or key words suggested？

I will be really appreciate for anything！！

[Abhishaike]

Hey! Honestly surprised that anybody checked out this repository, thanks for taking a look.

Would you be able to link me to the thread where I posted those comments? The context would give me a better idea of what research I was exactly referring to.

[Abhishaike]

Scratch that, found it! Concerning 'relationships between significant point': I was doing some research at a bioinformatics company and found that inputting some 'overall' information within the sliding window was very useful.

For example, let's say you were analyzing the QRS complex of an ECG signal, where each input to your CNN was the entirety of the wave. Instead of just throwing the entire raw wave into your CNN (which is what I did for the EEG project), it would be useful to engineer some features about that wave as static features. Such as the peak of the wave, the length of the wave, the R-R interval, and so on. Many of the useful features will be domain-specific, but you can brute force this engineering with a library like https://tsfresh.readthedocs.io/en/latest/ and throw out irrelevant features. I didn't do that this engineering process for this project because I was already so exhausted of having to deal with the data cleaning process that I just wanted to get this project over with 😅

As for any other advances, multi-task learning is something I feel quite strongly about and using that as a keyword would be helpful. Past that, just looking up '1D time series CNN' should get you pretty much any results you need.

Finally, let me know if you have any questions or want to bounce ideas off of me! I haven't gotten knee-deep in research since I started my job, so it'd be fun to go back to projects like this for a bit :)

[wstang35]

Thank you for your answer with patience!

I get the idea of auxiliary data now, which is interesting！ And It reminds me that I have a similar idea (I want to extract those points which amplitudes are greater than 75uV as an auxiliary input sequence to CNN, and this maybe helpful to my project, which is sleep stage classification using EEG), but I haven't implemented it.

However after studying your experience I think this maybe really useful and I will implement it ASAP.

Also, thank you for your keyword advice! I will go catch up some details and advances in 1D CNN, and see if I can use any of those to my project!

[Abhishaike]

Sounds good! Let me know if you have any further questions, hope the project goes well.