Ask HN: ML, EEG, metal 3D printer, evolved antennas = high resolution BCI? | Hacker News Reader

Ask HN: ML, EEG, metal 3D printer, evolved antennas = high resolution BCI?

2 pointsTomMarius9 comments

Hi, what do you think about the combination of these technologies?

EEG has a problem breaching the skull. Evolved antennas would be used to pick up better signal on various frequencies that are interesting. Because these need to be very small and ideally integrated into a head cap, we would use 3D printing to create these, which should also open room for even crazier (and more efficient) kind of fractal designs (fractal antennas are what phones use AFAIK). Then we would use GAN to clean up and another GAN to understand the signal. Because of (supposedly, that's what my question is mainly about) significantly better signal, it should have much better performance.

I'm not an expert in any of these. What do you think?

View on HN

I would not trust experiments using GAN on EEG to provide higher resolution/'quality' signals. You're not going to have a reasonable method to collect the data needed to train such a network, nor to be able to generalize the network. It will hallucinate patterns and given the dubious statistical quality of many comp-neuro papers I would be even more skeptical of many results.

If you're interested in applying some of these ML techniques I'd consider looking towards ECoG or LFP array recordings where sensor noise and sensor failure could in part be approached by techniques like this. For long term LFP arrays sensor failure is expected and models do not need to necessarily generalize to other individuals (and there is enough data to train on).

Overall you're not going to create signal when there was none in the first place.

I don't have the clear data but I know people that work on obtaining them and making them open (from local universities). I would like to use this method together with other methods, so even poorer performance than other methods could help - provide certainty in some uncertain edge cases, for example

The big question is whether the signal detectable on the outside of skull is closely correlated with what's going on inside

As an ex-researcher in this domain (still doing research, but elsewhere), the data does not currently exist. I believe you are grossly underestimating the amount of efforts, subjects, time, funds, and bureaucracy needed to obtain this data.

You are probably assuming I'm in the USA. I'm in Central Europe and this data is produced by local university (that has many great hospitals attached to it), open. I have friends who work on the project, they will start releasing data next year. They're working on the project for past 5 years.

Perhaps you should first learn what a GAN is, because what you wrote makes no sense.

I wrote the question precisely to learn more about these topics. Why do you think GAN doesn't make sense in that context? There are many papers about using GANs to clean radio signal, not this one specifically but as an example of a GAN/radio paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6263619/

In the paper you linked to a GAN is used to generate fake data. This fake data can help when training a classifier, just like any other data augmentation method.

I could not find any papers about using GANs to clean up radio signals. It could probably work if you have large datasets of both clean and noisy signals, but in this context you most likely don’t. And even if you did what makes you think denoising with GANs would work better than traditional methods?

Just because it’s a hot buzzword doesn’t mean you should use it.

Why are you condescending me? I studied applications of GAN nets for dozens of hours (still an amateur beginner of course). I didn't even know it was such a buzzword when I got started as I don't follow twitter etc.

Here is a paper that talks about exactly what I am proposing, using GANs, minus evolved antennas: https://arxiv.org/abs/1806.01875 (it is no secret that this paper has inspired me to write this question - my thoughts were about improving it with evolved antennas).

About the clean data requirement, it is common practice to record as much as possible once doctors open somebody's skull because such opportunity is scarce. So people are working on it.

I apologize, for some reason I was in confrontational mood.

This paper is also about using GANs to generate fake data.

Here’s the thing - to generate high quality fake data you need to have a lot of real data, but if you do then you don’t really need fake data.

Re denoising with GANs: do you have a lot of clean data (eg from inside the skull) that is equivalent to the noisy data (from outside the skull)? Let’s assume you do. Do you have any example of someone using GANs to clean those signals? Or any radio signals? What makes you think it would work better than any other existing denoising methods?

Reply on news.ycombinator.com