Hacker News Reader

Ayyyyy fellow neurosignal hater, I wasted the first year of my PhD (out of 3) working on electroencephalography (EEG) signals, I'm forever bitter.

You see all these great papers on medical applications but no one tells you it's a high-dimensional signal that contains one bit or maybe a byte of data at best. So you can do binary classification, or one of many classification (poorly).

And that's all you see in all studies:

Claim: "Learn to predict behaviour from EEGs"

Actual: "We learn to tell whether the patient is moving in any way or completely still with 60% accuracy."

Claim: "Learn to reconstruct images from EEGs"

Actual: "1 in n classification of previously seen images, then we put a billion parameters on top to go from the binary vector to the mean of the training images with no actual information used from the original signal."

>I'm forever bitter

Ahahaha, that sums it up 100%. And nothing will change as long as researchers can keep selling the same bullshit to grants bodies over and over.

There should be the ability to grant-hunt. Proof your opponents work is bullshit and get the money assigned to your project. Make hunting research false claims and grant hacks profitable for actual researchers and experts. For example proof irrelevance of research or previous work and get a percent of the grants.

Make it unprofitable and dangerous to be a fraud.

This is a really good idea. At my previous job I was handed a scientific paper to implement for an image classifying algorithm, hand picked by the team. Within a month the entire paper turned out to be fabrication, that's 160 man hours gone to waste. They still have the documents and the code as definite proof but it's worthless.

Do you agree that the 'consumer grade' devices that claim mind-control, and seem to work, actually respond to (facial) muscle movements - rendering them useless in (locked) in patients?

There are definitely external arrays that can work in locked in patients, but even when the stars align you'll usually be looking at something like 10-20 bits of data per minute.

For most people it'll be much less than that, and for some it won't work at all.

Generally speaking, these arrays are greatly outperformed by eye tracking, or even a nurse holding an A4 piece of paper and responding to blinks/grunts.

This in and of itself wouldn't be a problem, as the tech is in its infancy, but the real issue comes is that researchers are focused on bullshitting the public and selling science fiction rather than improving the existing technology.

Nobody understands how or why it works, and instead of learning about the brain we're just inventing new ML trickery.

I don't have enough experience with those to answer I'm afraid, I guess it'd depend on the number of electrodes and their placement.

The big problem with those recordings is the skull which acts as a low-pass filter over the signal that goes into your electrodes, more so than the equipment (within reason). So it could go both ways.

Yeah, I thought the same after seeing this. It's kind of a fun use-case of Diffusion models in this context, but as a scientific paper it seems too overselling. Well, it surely is the kind of clickbaity content to anticipate lots of retweets from.

I only skimmed the paper, but from what I understood, it is essentially a diffusion model pre-trained on a handful classes. The brain information is then used to largely "pick which class to generate a random image from".

The paper itself even picked the "better" examples. The supplemental materials show many more results, and many of them are just that, a randomly generated image of the same object class the person was seeing (or, the closest object class available in the training data).

"Reconstruct" seems a pretty bad word choice. I think the results are presented in a way vastly overselling what they actually do. But that's a worrisome trend in most of AI research recently, unfortunately.

(I have a PhD in a field of Applied Machine Learning. I work at a university in Computer Vision.)

VCs‘ due diligence is not “does it actually work” but “can it be pawned off to public coffers” (updating medicare “best practice” guidelines, inserting this tech into some gov procedure like maybe in criminal justice etc) so they might still be in the run, especially with bureaucrat-hypnotizing press-magnet “tech” like this. The money printer can keep up illusions of scientific progress for longer than we think.

> They're using ML to map the response from a test stimulus to some interpolated response from 50 training samples, arbitrarily assigning new sample images to these points in space and then passing it off as mind reading.

Could you clarify this? For example, if they trained the model on me visualizing a bear, a fish, and a bird, and then the neural net still outputs "horse" when I visualize a horse, I'd be impressed by this. And if they're going directly to images, I'll be more impressed if they get "black horse" and "brown horse" correctly. That's true even if the idea of reconstructing a specific image is kind of bullshit.

Is this what's happening? Or is their work just turning previously seen inputs into basically identical outputs?

> if they trained the model on me visualizing a bear, a fish, and a bird, and then the neural net still outputs "horse" when I visualize a horse

Well, the failure cases figure says it does not work if "training and testing datasets do not overlap". So, it'd just find the closest trained class and then generate new image from that class (i.e., in your example the bear might look more similar to horse than a fish or bird, so it'll generate a random bear).

I don't think it quite says that. Exact quote:

> We assume the failure cases are related to two reasons. On one hand, the GOD training set and testing set have no overlapping classes. That is to say, the model could learn the geometric information from the training but cannot infer unseen classes in the testing set

Now, if all the GT results had been fails, it might be reasonable to conclude that it doesn't work if the sets don't overlap. However, there are only 6 that they graded as fails. (A few more look iffy to me.) If I'm reading their statement correctly, there was no overlap between the two sets:

> This dataset consists of 1250 natural images from 200 distinct classes from ImageNet, where 1200 images are used for training. The remaining 50 images from classes not present in the training data are used for testing

And if I'm understanding this correctly, that makes the results look sort of impressive. I mean, at the very least, the model is getting the right class from the testing set most of the time, even though that class wasn't in the training set. That's ... not ... nothing?

On the other hand it seems they cherry picked the best of five subjects for the results they show in the supplementary, which is ridiculous.

> Subject 3 has a significantly higher SNR than the others. A higher SNR leads to better performance in our experiments, which has also been shown in various literature.