I'm highly skeptical. I mean, a hash function that has four output states also maps anything to one of those four states. That doesn't mean it's some next-level classifier.
The problem here is EEG. EEG bandwidth is not enough to capture that much information. There is far too much noise introduced by the skull and muscles. It's most likely physically impossible to do something like this with EEG.
What's likely happening here is that there's some large scale oscillations that are sufficiently unique to discern the images from each other. This does not mean they are reproducing the images. I am highly skeptical of the methods used here -- they are almost certainly flawed.
I, too, once had dreams of conquering the planet with EEG when I was a grad student. I quickly learned that physics makes this infeasible. Anyone who is serious about BMIs are studying invasive BMIs and how to make them as safe as possible. Going inside the brain is unavoidable, I'm afraid.
About a decade ago when I was still in school, I did some work in brain machine interfaces as well as a friend. I made an EEG from scratch, worked on the DSP and amplifications to make it all work, and also had access to a much more expensive state-of-the-art machine. While I didn't work directly on the project with my friend, at the time they came to the conclusion that non-invasive neural processing (so something topical like an EEG, no surgical implants) could process about 1 bit per second of useful information - the noise to signal ratio was about 1000:1. When most people read the raw data from an EEG they don't realize they can't even see the actual data - they're seeing eye movements, facial muscle twitches, and other noise artifacts that overwhelm the actual signal. I'm guessing the technology has improved a lot since then (I'm in another field now), but it's hard to imagine it gaining however many orders of magnitude in resolution necessary for this to be viable.
If you have multiple different points where you measure, which all have this overlapping signal problems but at different strengths, couldn't you hypothetically build up a model that "solves" these different weights and untangles the signals?
Yes, but... At the end, you're still reconstructing pieces of information from something that was almost destroyed. Picture it this way: there are amazing deconvolution algorithms that can "undo" all sortf of noise and lack of focuse -- but the end result, however good to the original "bad" data isn't nearly as good as a well taken image to begin with.
Disclaimer: I work in image processing, so the example may be a bit obvious to me.
Yes, that'd be a better analogy. My point was that, even if you had the best reconstruction in the world, having to reconstruct from a degraded source is worse than working from a good source to begin with.
From a practical perspective they don’t have to be perfect or as food as the original, not even close, just good enough. “Good enough” though is also extremely hard to achieve, assuming it’s even possible with this technique.
There is at least one 'affordable' fNIRS device coming to market that looks promising, https://foc.us/fnirs-sensor/ There's a paper somewhere on using machine learning to help identify signal, this one is specifically about pain, https://www.nature.com/articles/s41598-019-42098-w Say for example you were making an insurance claim for neuropathic pain, this kind of information could be very important.
Example of how it may be overfit: Waterfalls are loud, audio regions of our brain may activate in response to waterfalls. Classifier reads that to predict waterfalls.
Great example! Similar thing probably holds for moving main limbs. A good EEG should be able to pick up when you think about moving your hand or leg. I doubt a good EEG could distinguish more than a few dozen patterns. Most experiments have trouble with even a handful of patterns. Still could be useful, but just very limited.
Watch how it has preconceived notions of these scenes. It frequently fails to reconstruct the correct scene from video, and it also turns completely blank input into one of the scenes it was trained on.
This will be disastrous on so many levels. The eyewitness might not have a good sample of a minority race. The GAN dataset itself might also only be trained on celebrity faces so it doesn't know how to generate anything else (e.g., a teen).
But it will be deceptively high resolution so police will rely on it.
OR they will just find out that this method can just as easily be producing a picture that someone who is good at visualizing just made up in their mind and is actually "looking at" in their mind's eye. This making this technique useless as a form of truth seeking machine.
as someone in this thread pointed out, the model is overfit.
I did my thesis on EEG signals, also having a very idealistic view of what could I do with it, only to find that even the most basic of tasks is hard to classify (even to find motor cortex movement intention signatures (if you want to move left or right hand)).
This work will not go into the real world in this stage, as it is badly done and most certainly having multiple flaws in the implementation, rendering it unusable in the real world.
So, don't get too stressed out about this, if it happens, it will be about 20-30 years from now. And keeping in mind how slow the law enforcement technology moves forward (aren't most of them still using windows xp and vista?) I would count more like 30-50 years.
My research group is doing the same thing but with music. Music may be more promising than images because of the Frequency Following Response -- a sort of direct resonance effect in the brain in response to sound.
We have 24 subjects listening to 12 songs in random order, with 128 channel EEG sampling at 1000hz. We can then label all these data points with the musical features at the time the data is collected.
We don't have a public repo yet, but we are sharing data.
I dont think their model is working, and Im not sure it ever will. Simply reading brainwaves, a bi-product as I understand it, of the actual neuron activity, couldn't possibly give you an accurate result.
The end results are much, much better than I thought they would be. Luckily, I think it would be easy to fool the training by thinking about a totally different image to the baseline one. Idk if that would stand up to rubber hose cryptanalysis, but there’s got to be a way that can.
The end results show an overfit model. It's not predicting that specific input out of an option space of everything; it's essentially predicting that mode (out of the 4) and probably capturing things like "if brain's audio regions are active, it's a waterfall, because waterfalls are loud and trigger that".
I read in some article some time about use of SQUIDs (https://en.wikipedia.org/wiki/SQUID) to map activity of a single neuron non-invasively. There was a lot of hype at the time for brain-computer interfaces based on that, but then same as with many technologies that were "just 5 years away" those 5 years came and went with no deliverables expected.
There’s a great movie called Until The End Of The World that centers on this kind of technology. Once the scientists get it to work, they realize that they can record and play back their dreams, and they become addicted to watching them.
Lena source image resulting in some other random woman "reconstruction" = model over fitted AF. Put a dead fist and it will continue generating "reconstructions".
Dystopian because these are very high tech which naturally concentrates their use and abuse by the super rich nation states and corporations while the rest of humanity will be watched like caged animals with no scope for any kind of opt out or pushback.
The problem here is EEG. EEG bandwidth is not enough to capture that much information. There is far too much noise introduced by the skull and muscles. It's most likely physically impossible to do something like this with EEG.
What's likely happening here is that there's some large scale oscillations that are sufficiently unique to discern the images from each other. This does not mean they are reproducing the images. I am highly skeptical of the methods used here -- they are almost certainly flawed.
I, too, once had dreams of conquering the planet with EEG when I was a grad student. I quickly learned that physics makes this infeasible. Anyone who is serious about BMIs are studying invasive BMIs and how to make them as safe as possible. Going inside the brain is unavoidable, I'm afraid.