Researchers Use AI to Generate Images Based on People's Brain Activity | Hacker News Reader

Researchers Use AI to Generate Images Based on People's Brain Activity

vice.com

54 pointsisaacfrond26 comments

Dupe: https://news.ycombinator.com/item?id=35012981

I'm pretty sure this kind of thing has been done before, at least very lofi images. Maybe not in "high-resolution", but I really question the scientific value of doing so.. it must necessarily be hallucinating details, similar to the common criticisms of upsampling networks. There is no way we are reading brain activity at sufficient details to reconstruct a "true" image directly from the data without a lot of inductive bias. There isn't even evidence that such data exists. (that at the lowest level brain activity actually corresponds to image-related detailed information). The most we have afaik is proof that certain neurons in the optical pathways respond to certain types of lines. I'm interested in being corrected on this point. But roughly inverting some relationship with random seeds of a generative network sounds like pretty iffy methodology to me. At most it's probably categorizing the brain activity and generating a random image based on that. (still cool of course! I'm just doubting that the high resolution images contain information that really reflects brain activity, that fact that this is possible is still a neat trick)

I'm basing this criticism on the linked Vice article though, haven't read the paper yet admittedly.

I see a resemblance between the question whether this AI is truly decoding brain activity into images, and Scott Aaronson's philosophical waxing about whether waterfall truly compute functions (answer: no, because the algorithmic complexity of the necessary decoder is at least equal to the complexity of the problem being computed)..

People on Twitter are jumping to the conclusion that this allows you to read someone’s mind but maybe all it does is allow to you to reconstruct something similar to what they are looking at.

For example if you recorded the activity of every cell in the retina, you would be able to do good image reconstruction but no one would interpret this as mind reading.

It’s possible that this is a similar story but instead of the retina, the decoder is just getting all of its signal from early visual cortex and therefore it’s not doing anything that a well placed camera pointing the same way as the subjects eye couldn’t do.

Yes, in principle it’s possibly doing more than that and using signal from eg inferotemporal cortex, but how much? There is no evidence it is doing anything beyond simply leveraging the retinotopic organization of primary visual cortex here.

They say published, and link to a pre-print. It's not peer reviewed.

Not yet. The paper looks legit to me. Publishing first on a pre-print archive is pretty standard precise.

https://www.biorxiv.org/content/10.1101/2022.11.18.517004v2....

It's also a great way to spread misinformation if the paper has issues. This post goes viral, but news about it getting retracted won't. The article shouldn't refer to it as "published in December". It should say that the paper is not peer reviewed yet.

That's 'tech journalism' for you btw, especially when it comes from cesspool quality writers from the site at the OP.

This is truly dystopian. One step closer to the thought police

Or dreaming thoughts into reality. Paintbrush of the mind making imagination real.

This reconstructs images from a specific person /while they were looking at the original image/. It's not mind reading, it's eye reading. It's not known to work across people either.

But you can imagine a later iteration doing that?

If they could read something other than a cooperating subject in an MRI machine, sure.

But it's slowly getting there. Researchers will keep trying to infer as much of your cognitive processes as possible. This is a small step towards that, and not even the first one (which is over 10 years old, I think), but should not be dismissed so lightly. It won't go fast, because every step is a small one, but at one point you're going to have to say "Well, it's too late now. There's no going back anymore. Sorry!"

I’m not sure it’s feasible to stop things like this.

You can try to implement global prohibitions, but our species has a pretty bad track record at that unless the thing we are prohibiting is a stable equilibrium in game theory (anybody who can pay to play wants to stay in the equilibrium).

AI is still expensive, but it’s not so expensive that only megacorps and nation states can pay-to-play. To fully stop this kind of development, you’ll need to convince everyone with the means in every society to stop tinkering with this stuff.

I personally think we’ve crossed the threshold where inspired hackers in their basements can build nifty things with this with or without your sanction. You might be able to slow progress, but there will still be a black market of knowledge and implementations that continues pushing this forward. Not just because it’s profitable, but because it’s fascinating too.

What kind of failures are you thinking of? Full scale human cloning for example hasn’t happened even though it’s been possible for decades.

Chlorofluorocarbon (CFC’s) are basically gone. Blinding lasers have actually been banned etc.

I’m honestly having trouble thinking of failures here.

Nuclear (mostly) requires megacorp or nation state funding and most players are in an equilibrium for now. Supply chains for nuclear are feasible to detect and control by existing players who want to keep the equilibrium.

When you look at markets with relatively low barriers to entry compared to nuclear with correspondingly high recreational and/or market value - humans have faired far worse at effective prohibitions.

Alcohol and other drugs. Illegal porn. Explosives. Sharing copyrighted material. Guns. Buying and selling humans. Cryptography. Etc.

Hop on any Silk Road replacement and you’ll find a nice list of failing prohibitions.

I do believe AI falls into the bucket of having a feasible black market. The supply chains are hard to detect and control. The cost is relatively low. And both the recreational and market value of it is high. It seems like something humans are going to do at this point regardless of whether it’s sanctioned or not.

Ahh, I see what you mean but none of that takes heavy R&D.

Suggesting minimal barriers to entry seems like wishful thinking at this point. Sure, the average CS student should be able to write a pathfinding algorithm etc, but people who want to restrict types of AI research have a higher bar in mind.

Put a moratorium on publications. It cuts off the oxygen. Most research is done for publications. You need them for tenure and grants. And it stops propagation.

This result is not meaningfully different from things that have been done with MRI 10 or so years ago. Why do you think it’s getting there? I see this field not moving forward in any meaningful way.

Try it with a person with Aphantasia.

Could definitely see this being used instead of lie detector tests/interrogation. Obviously not perfect but the utility def seems there

Also not news, posts quality is going down, sorry for redudant and useless post, well not as useless as the post. Plus, vice? really?

I wonder if they'll ever be able to do it with EEG/facial EMG/Accelerometry or some other wearable consumer grade tech?

Although people have used EEG as game controller. So getting a few bits worth of information out of EEG is definitely possible.

But I doubt that EEG contains enough information to encode an image in. So, just from the amount of bits needed, I don't think that it is even possible.

Extraordinary claims….

Reply on news.ycombinator.com