You Cannot Reliably Reconstruct What Someone Saw or Thought From Only an fMRI Brain Scan

Another day, another unfounded claim of "mind reading from brain scans." 

Brains do not have any representation of what a person is thinking or seeing. And the idea that you could ever detect what a person was thinking or seeing from an fMRI scan is unbelievable. What is it that an fMRI machine actually detects? Speaking roughly, we can say that an fMRI machine detects changes in blood flow in different parts of the brain. Speaking more precisely, we can say that the exact thing measured is differences in hemoglobin levels in different parts of the brain. Hemoglobin is a molecule that contains iron. And machines such as fMRI machines are all based on magnetism, magnetism that can detect iron levels very precisely. 

This description may surprise a typical person who has been led to think that fMRI machines detect changes in electrical activity in parts of the brain. An fMRI machine does not directly measure changes in electrical activity in the brain. All that it measures are changes in blood flow, and hemoglobin changes related to changes in blood flow. The Google Gemini graphic below tells you precisely what is measured by an fMRI scanner. 

So how should we react when we see an article in the press like the recent one entitled "This ‘Mind-Reading’ AI Is a Wiz at Figuring Out What You See"?  The article has this paragraph:

"Brain-IT takes advantage of the thing that AI is best at, which is pattern recognition. It takes in data from functional MRI brain scans, which track changes in blood flow and oxygen in the brain to measure brain activity, and then uses that data to recreate the images the person was thinking about when the scans took place."

As soon as we read that paragraph, we should say to ourselves "this must be a shady trick." That's because what a person sees and thinks about sure as heck is not represented in the brain as changes in blood flow.  

The quote above does not even match the title, which refers to what people were seeing, not what "the person was thinking about." 

When you encounter cannot-possibly-be-true claims like this made by neuroscientists or their lackeys,  it is usually pretty much a waste of time to try to track down exactly how the ruse was carried out. Neuroscientists who engage in these kind of misleading parlor trick shenanigans are pretty good about "covering their tracks." Typically, they present their results in papers that discuss some extremely convoluted and opaque methodology, one that makes it hard to untangle the funny business cheats that were going on. It's rather like some money launderer transferring funds between shell companies to minimize the chance that some investigator will be able to track down how the money was laundered. 

But there are some things you can do to expose the baloney that is occurring. One is to simply look for whether the paper authors ever claim that their not-really-mind-reading results were obtained solely from inputs of brain scans. If you fail to find such a claim, it is a pretty sure sign that the results were not obtained solely from inputs of brain scans. 

What typically goes on is some situation rather like this:

Inputs:

Brain scans

Something else

Some other thing (maybe some sneaky data back door)

Some other thing not mentioned in the press release

Outputs:

Some image looking a little like what someone saw or thought of

The result will be sold as an impressive result produced "from brain scans," without any mention of the crucial inputs that were not brain scans. And typically we will have displayed some best-looking results, and we should ask: were these cherry-picked best results, or were they the average results produced? Typically we don't know what is the answer to that question. It's easy to get best-looking results if you cherry-pick best results. For example, if you spend all day guessing the results from a random number generator generating three digits, you'll probably be able to make an impressive-looking table of good guesses, if you are allowed to cherry-pick best results. 

A look at the scientific paper makes us suspect that the results are not so impressive. The paper misleads us in the first sentence of its abstract, claiming, "Reconstructing images seen by people from their fMRI brain recordings provides a non-invasive window into the human brain."  You cannot do any such thing using inputs of only fMRI brain scans. 

It takes some practice to be able to find the part of a paper like this where the authors reveal that their outputs were not produced solely from inputs of fMRI scans. Typically the authors of paper like this write paragraph after paragraph of the most convoluted prose, attempting to "cover their tracks" so that it is not easy to discover their cheats. In this case I found in the BRAIN-IT paper this paragraph, which lets you know the shady business that was occurring (I will underline and boldface a phrase not underlined and boldfaced in the original text):

"Enriching the Training Data. The training is performed using pairs of fMRI measurements and the corresponding images viewed by the subject, taken from the NSD dataset (see Sec. 5.1). The model is trained on data from all subjects simultaneously. Since the subject-specific data is limited, we further enrich the training set with 'external images' (natural images without any fMRI). Specifically, we use ∼120k natural images from the unlabeled portion of COCO dataset."

The Section 5.1 referred to says this:

"We used the Natural Scenes Dataset (NSD) (Allen et al., 2022), a large publicly available 7-Tesla fMRI dataset that records fMRI of 8 subjects as they viewed diverse images drawn from COCO (Lin et al., 2014). The dataset contains ∼73,000 Image-fMRI pairs, comprising ∼9,000 unique images per subject and 1,000 images shared by all subjects."

So what is going on seems to be something like this:

Inputs:

1. fMRI brain scans taken when people saw particular images.

2. The exact images that those people saw during such brain scans.

Outputs

Something sometimes looking pretty close to what was seen by those people when the brain scans were taken.

The paper has the title "BRAIN-IT: IMAGE RECONSTRUCTION FROM FMRI VIA BRAIN-INTERACTION TRANSFORMER." It seems the paper should have been titled something like "Reconstructing Images from Brain Scans of People Seeing the Images and Also From the Images They Saw." If the paper had that title, people would know that nothing of much significance was going on. 

You cannot reliably reconstruct what someone saw or thought from a mere fMRI brain scan, which does not even directly measure brain activity or brain states. So whenever you read a paper like this in the future, you can be pretty confident that some kind of cheat or parlor trick or shady business is going on. It's almost always some kind of data back door cheat, whereby the paper authors smuggle in non-neural inputs, often inputs that are pretty much the same as the outputs. 

A press release on this paper seems to have an untrue caption at the top an image entitled "Pictures on the Brain." We see a grid showing a set of image pairs, and each pair looks very similar. We read the claim that the images on the left are images shown to people who were brain scanned, and that the images on the right are the images produced by the "Brain-IT" program or model, and that this model or program had "never before seen" these images or the corresponding brain activity scans. This does not match the statements in the scientific paper, which makes no corresponding claims. The scientific paper has  a grid image (Figure 5) similar to the "Pictures on the Brain" image in the press release, but with different images; and the caption of Figure 5 of the paper does not claim that the images on the left are images shown to people who were brain scanned, and that the images on the right are the images produced by the "Brain-IT" program or model, and that the model had never seen before these images or the corresponding brain activity scans. The scientific paper also has lot of other similar-looking image grids. But the captions of those grids do not claim that the model had never seen before these images or the corresponding brain activity scans.

The press release uses the phrase "never before seen" as described above, but the paper never uses the word "never" in its text, and does not have any relevant corresponding phrase meaning "never before seen." My best guess is that we have here some anonymous institutional press release writer doing what institutional press release writers so very often do these days -- making false or misleading statements about the research the press release is promoting, while trying to make it look like some great feat was accomplished when that was not done. 

The scientific paper has lots of impressive-looking image grids comparing one image on the left and an almost-identical-looking image on the right (they look like the top half of the image below). But the captions of these image grids are so short and skimpy (consisting of only a sentence or two) that readers cannot tell exactly how the results were obtained, and we can have no high confidence that anything very impressive has occurred.  If the authors wish to command attention for their results, they should replace their super-skimpy figure captions with full, detailed captions giving maybe 15 or 20 sentences describing exactly what went on to produce these image grids. 

The paper "Spurious reconstruction from brain activity" refers to problems with some papers using the same Natural Scenes Dataset mentioned above. The paper says, "In the field of machine learning, 'zero-shot' prediction refers to the ability of a model to accurately predict or classify novel contents not encountered during the training phase (Larochelle et al., 2008, Palatucci et al., 2009). " In the paper for this Brain-IT model, we do not hear the phrase "zero-shot" mentioned. So why was a caption of the press release discussed above making it sound like such "zero-shot" prediction had been achieved?

The paper suggests that there is something about this Natural Scenes Dataset that makes it particularly easy to get impressive-looking  results when using it. Below is Figure 2 from the "Spurious reconstruction from brain activity"  paper.  We see that some models get impressive results using the Natural Scenes Dataset. But when a different image dataset is used (the Deeprecon image dataset), the models perform poorly. 

There is one other important consideration. When we see some paper showing some "successful image reconstruction grid" like those we see in the Brain-IT paper discussed above, we should always ask: were these successful-looking results typical results from the model trying to reconstruct what someone saw when his brain was scanned, or instead were the displayed results some best results cherry-picked from a much larger set in which results that good were very untypical? I doubt that we get the answer to that question from either the press release or the the Brain-IT paper discussed above.

The Brain-IT paper discussed above has not yet been published in a scientific journal. At the end of the paper the authors say, "Upon publication, we will release our full code with saved checkpoints to ensure that all results are easily reproducible." Inspection of the software code in projects such as this will often reveal various types of funny business, cheats or machinations. 

Previous efforts of this type involved the most outrageous types of cheats and shady sleight-of-hand. In my 2023 post here I describe one effort. Experimenters grabbed text annotations from an image database (descriptions of what the images showed), and then fed those text phrases into an AI image generator, to produce images that were then claimed as images "derived from brain scans," because brain scans were included in some convoluted pipeline that included such a technique.  That was something that was the shadiest smoke-and-mirrors trickery, something so bad any decent scholar should have been ashamed of such work. 

I don't know whether trickery that outrageous is going on in the latest claim of "mind reading from brain scans." I do know this: in general, what someone saw or thought cannot be reconstructed from mere fMRI brain scans taken when that person saw or thought something (largely because such scans merely detect blood flow changes in the brain). So whenever anyone suggests or insinuates that what someone saw or thought was reconstructed from only fMRI brain scans, there must as a general rule be some bad kind of trickery or badly misleading language or very misleading funny business going on.