Hacker News Reader

Comment by ad404b8a372f2b9 | Hacker News Reader

Not much to see there.

I don't have much faith that the stated goal is realizable, I don't think fMRI data holds enough information for faithful reconstruction. What it will end up being, like all projects like it before, is a massive network that contains:

- An implicit multi-class classifier that returns 1 out of n previously seen images

- Gigantic model on top to hallucinate a reconstruction close to that previous image

In previous projects of this type the reconstruction was blurry because it was essentially the mean of the previously seen images close to the label predicted by the implicit classifier, in this project I imagine they will be photo-realistic since it's a stability.ai project but no closer to the content the subject is thinking about.

View on HN

I don’t know. Now that you explain how it works, it’s starting to worry me. Imagine being hooked up to an ML model designed to classify whether you’re thinking about porn or not, and emits a beep whenever it detects it. The implications of that are astonishing, if it’s accurate.

It hinges on accuracy. If it can detect a topic, the image reconstruction is almost moot, because you can start interrogating people about what they know.

I wish I knew more about biology to know whether this is at all feasible, or if it’s just smoke and mirrors. (A distressing number of projects like this are to secure further funding for a lab, not really to advance science.)

Since you seem to know about this, do you have any details about the classification process? I assume it requires the cooperation of the user, but it’s easy to imagine that a “primal brainwave” pattern might be common to all humans.

> classify whether you’re thinking about porn or not, and emits a beep whenever it detects it

Already gone down this road with lower tech. In the 50s - 70s it was in vogue to monitor blood flow to the genitals while subjecting the subject to erotic imagery of various kinds to see what turned them on. [1]

It was then realized that arousal is also correlated with pupil dilation, which meant that a much less invasive test was possible. Which of course led to the Canadian government (and probably other governments) using it in the 1960s to detect possible homosexuals in the public service. [2]

The scientific flaws and human rights violations of such a test are too many to even begin to list in a short post like this. Still, the inconvenient and disturbing thing is: it does work. Sort of. Kind of. Not good enough to be confident about the result with an individual, but a properly run protocol has better than chance odds at determining if a man is gay or straight, in terms of aligning with their self-reports about what arouses them.

Whether someone is actually erotically aroused by something based on a physical response is obviously controversial. For example, a repeatable finding is that self-reported-heterosexual men with intensely homophobic views are more aroused by homosexual male imagery than the average heterosexual male. Commonly reported in the news and media as: homophobic men are secretly gay. And I guess that's one possible interpretation. But someone about to to fly into a fit of rage is also aroused. Sometimes literally -- sexual arousal during flight-or-fight and disgust responses is a thing. So it's tricky to interpret what that result means, if anything.

No doubt we will fall into similar traps again, with the brain scanning version of it.

[1] https://en.wikipedia.org/wiki/Penile_plethysmography

[2] https://en.wikipedia.org/wiki/Fruit_machine_(homosexuality_t...

I’m struggling to find a reference, but I’ve already seen this used in the reverse during interrogations.

You use something like an EEG and show someone a series of images to establish a baseline of “familiar” and “unfamiliar.” Then you show them photographs of the crime scene that haven’t been made public to test for familiarity.

I question its accuracy but your scenario already seems possible.

Edit: for siblings comments about journalists - “identifying” a source could probably be “accomplished” using the above technique already, no need for AI. Show photos of random people mixed in with the photo of the suspected leak in your gov. department.

lol yes, the social implications are horrible for this kind of work but I'd say more so for dissidents, journalists, and so on rather than porn.

Projects that are at the intersection of life sciences and machine learning are rarely interesting along both axes, it's always novel on one side and the other side is just there to make it seem cool. For example you'll find a lot of ML scientists investigating medical issues but lacking the basic medical knowledge to actually make an impact, similarly you'll see doctors using neural networks where a linear regression would do. In both cases the paper seems novel to the people in the field it's being published in.

The classifier I was talking about is implicit, if you look at their proposed diagram they want to use fMRI embeddings, they serve as a continuous equivalent. You put a sufficiently large model on top of any embeddings and it will learn to reconstruct their modes. The embeddings are trained across subjects so it's not specific to a single patient but they hope to fine-tune it per patient using latent space alignment.

If you look at the papers in their recommended reading list, it's apparent that only low-frequency signals are learned, for example for face reconstruction the model can learn hair color, sex and beard/no-beard, but the actual face doesn't look anything like that of the person being looked at. Similarly for other benchmarks, the model can tell it's looking at a plane, but not much beyond that. This is cool from a technological standpoint but I recall papers doing the same thing at least 7 years ago. The novelty of their approach will be the architecture that comes after the embeddings, which can be interesting but won't add any more information to the fMRI scans.

> it’s starting to worry me

Not any time soon: https://duckduckgo.com/?q=fMRI+device&ia=images&iax=images

I'm actually struggling to find the usefulness of this. The person needs to be in an fMRI machine, so we're not talking about some sort of unobtrusive scanner in a public area. On top of that, in cases like what you're describing which basically just classify thoughts to a binary level, then it's pretty much useless because if someone knows that they're being scanned to see if they're actively thinking about pink bunny slippers, then they're going to think about pink bunny slippers.

> Imagine being hooked up to an ML model designed to classify whether you’re thinking about porn or not

Or whether you're thinking thoughts the government doesn't like.

There's already been preprints released showing fMRI reconstructions that appear to do better than an implicit multi-class classifier [1] [2]. But also, even if the result is an implicit multi-class classifier, if the n is sufficiently high then that would still be quite impressive!

[1] https://openreview.net/pdf?id=pHdiaqgh_nf

[2] https://arxiv.org/pdf/2211.06956.pdf

I see no evidence that they do better than multi-class classification, in fact they both work as I described. They learn embeddings of fMRI which perform an implicit classification of the data (which can be recovered just by quantizing the embedding space to get its modes) and put very large generation models on top.

The only reason the reconstructions are much better than before is because they use the latest generation models. Those models have internal models of the classes which allow them to fill in the high-frequency details in the reconstruction. The only information they get from the fMRI is the same low-frequency signal that previous papers already had, and indeed the only things the reconstructions get right are low-frequency: class of object/scene, broad position of object, broad shape of object.

fMRI scans are aggregates of brain information, they act like low-pass filters over the brain state. You can put as big a model on top as you want it won't make it more truthful a reconstruction.

I think, as you say, that detecting as many classes as possible is already a pretty good goal, developing new embeddings and techniques to see how much juice we can squeeze out of the scans. I like the arxiv preprint you posted in particular since it does just that and evaluates accuracy (although the way it does it is flawed since it uses an image classifier on the reconstruction which presents the same problems). What I don't like is the misrepresentation of what's going on when people put those large generative models on top of this kind of data.

I generally agree with the assessment that the limiting factor in quality is the signal from the MRI. Just one slight correction to the description of previous projects. We first built models that predict the voxel responses from images using ML. Then recorded the fMRI response on a new image. Then we ran many many images (not just those previously seen that have an fMRI response) through the voxel models and picked the top N images with predicted responses close to the actual response. Basically we used the big list of images as an approximate "natural image prior". Then we'd often show the average of that top N images so you get one picture (which destroys interesting multi-model stuff, but we didn't have good stable diffusion models of images back then that could do a gradient search on or something).

If you wanna see what the reconstructions look like without fancy image priors (to get a better idea of how little signal there is) check out this figure from one of the early papers: https://pubmed.ncbi.nlm.nih.gov/19778517/#&gid=article-figur...

Reply on news.ycombinator.com