> This breakthrough supports the theory that significant sex differences in brain organization exist, challenging long-standing controversies.
It doesn't, because there is no explanation for how it determines the difference. It's just as likely some confounding thing that's not causally based on the brain, like all these other "AI can tell" studies that don't identify an actual mechanism.
That said I'm surprised that we can't tell men and women apart from the brain. I would have assumed there would be some differences that would make it obvious at least for the average brain. It's more surprising to find they are indistinguishable.
>It doesn't, because there is no explanation for how it determines the difference.
You don't need to know the mechanism to prove that it "supports the theory that there is a difference in brain organization". If there wasn't a difference, it wouldn't be able to find it.
>It's just as likely some confounding thing that's not causally based on the brain
Given that it does it's identification by looking at the brain, whatever the cause is, it is ALSO manifested as a "difference in brain organization".
>That said I'm surprised that we can't tell men and women apart from the brain.
Who said we can't? The article points to that we can.
And there are other brain attributes besides organization we can use to tell them apart, the biggest being size, including size of certain structures.
> Given that it does it's identification by looking at the brain, whatever the cause is, it is ALSO manifested as a "difference in brain organization".
I think the person you’re responding to is talking about confounding factors _outside_ the brain scan, like when a model that could identify skin cancer used the presence of a ruler as its most weighted input: https://venturebeat.com/business/when-ai-flags-the-ruler-not...
This is a huge problem in medical use for AI - you have non-technical medical professionals influencing the data used to train the model, and introducing biases in the data that the model is learning from.
>I think the person you’re responding to is talking about confounding factors _outside_ the brain scan, like when a model that could identify skin cancer used the presence of a ruler as its most weighted input
Well, those are just "bugs" and the results are then useless.
My point is that results taking into account only the scan (not random correlations with outside factors), would indeed be able to prove differences between M&F brains, without needing to also give a mechanism for it.
Some silly examples of stuff it might find that aren't actually brain differences:
1. There's an M/F letter or other identifying text on the scans
2. Men and women get put into the machine slightly differently, and the ai sees a slight difference in the pose of the scan
3. It sees something else that's identifying and pretty good at splitting, like ear holes for ear rings
I remember reading about one of the skin cancer finding ML models that was developed. Most of the pictures from malignant moles were taken from a specific office and the algorithm was picking up on the lighting in that room rather than other differences in the mole to determine likelihood of it being cancerous. It did a great job of identifying photos based on the type of lighting used, but was overall less effective at predicting cancer.
I’m reminded of the tale of a similar breakthrough, and upon ablation, it was found to be indexing on a particular artifact found on one X-ray machine. Or maybe the patient name?
Anyways, would recommend some caution, there’s no reason to believe there’s some obvious fundamental difference in brain organization that we can’t see ourselves, until the research is a bit further along :)
Unfortunately, these things can be extremely subtle. Far too often it boils down to sampling bias where the differences are actually differences in the variance estimation that results in a significant result. For example suppose that "male" dice are biased to role a 5 slightly more often than "female" dice. You will find a statistically significant difference related to the 5-side of the dice if you roll them enough. It doesn't mean that rolling a 5 indicates anything about the "gender" of a particular dice.
> It doesn't, because there is no explanation for how it determines the difference
I don’t think an explanation is required for this.
I should be able to state that two species have a different lifespan based on a difference in mean age of death - without needing to provide any specific biological mechanism in my study.
Substitute that mean for a linear regression / neural network and the species lifespan for some vector of brain organisation to get the results in the article.
I was thinking about this too, but I don’t know how widely they overlap? Like, if (say) 30% of male brains are at least 10% smaller than the average male brain, and 30% of female brains are at least 10% larger than the average female brain, then a model that just considered volume might not be very accurate at all, because anything in the middle still has a significant chance of being either sex?
> I should be able to state that two species have a different lifespan based on a difference in mean age of death - without needing to provide any specific biological mechanism in my study.
You need to control for confounders. For example, if I told you:
- Mean lifespan of species 1 is: 3 years
- Mean lifespan of species 2 is: 10 years
Does that mean that species 2 has a longer lifespan than species 1?
What happens when you find out that species 1 is domesticated cattle, which has (almost) all of its males slaughtered at age 2?
You could then say "I meant for two species that are in controlled conditions", but now you need to define what conditions you're controlling, which implies what casual mechanisms you're controlling for.
> which implies what casual mechanisms you're controlling for.
Without thinking terribly deeply about it, it seems like this could be modified to
"what correlated mechanisms we're controlling for", in which case, yes I think if we controlled for the highly correlated variable of "does this cow live on a meat farm" with the mean age, we would be able to make broad claims about cattle lifespan without needing to be privy to the internal mechanisms of the agriculture industry
> Until recently, a model like the one Menon’s team employed would help researchers sort brains into different groups but wouldn’t provide information about how the sorting happened. Today, however, researchers have access to a tool called “explainable AI,” which can sift through vast amounts of data to explain how a model’s decisions are made.
> Using explainable AI, Menon and his team identified the brain networks that were most important to the model’s judgment of whether a brain scan came from a man or a woman. They found the model was most often looking to the default mode network, striatum, and the limbic network to make the call.
What's more fascinating it that it's not just a parlor trick and actually can be "validated" in some sense by doing a cognition test for just that part of the brain and observing the differences. So the model said "check here and you'll see a difference," and apparently, they did.
> They developed sex-specific models of cognitive abilities: One model effectively predicted cognitive performance in men but not women, and another in women but not men. The findings indicate that functional brain characteristics varying between sexes have significant behavioral implications.
> “These models worked really well because we successfully separated brain patterns between sexes,” Menon said. “That tells me that overlooking sex differences in brain organization could lead us to miss key factors underlying neuropsychiatric disorders.”
FWIW, feature maps are not really reliable (to put it mildly). Knowing what area the model is "looking at" is virtually meaningless if you don't know what it's looking at. We already knew it was looking at the brain so is it turtles all the way down? Those maps are great for confirming what people want to see "oh look it tells a cat from a dog by the eyes" but have no explanatory power. If they did then the paper would be about AI finding a physiological difference and would tell us what it is, not some handwaving about "here's where it looked".
> Knowing what area the model is "looking at" is virtually meaningless if you don't know what it's looking at. We already knew it was looking at the brain so is it turtles all the way down?
Huh?
I’m not sure what you are saying here.
If there’s some subset of the pixels or whatever which the models predictions depend on more than the others, then, it depends on those more than it depends on the others.
Obviously it would be using the data it is given. That, doesn’t mean that looking into “what part of the data that it is given, is it using?” not meaningful.
> “These models worked really well because we successfully separated brain patterns between sexes,” Menon said. “That tells me that overlooking sex differences in brain organization could lead us to miss key factors underlying neuropsychiatric disorders.”
Although never demonstrated, it's always been most intuitive to me that sex-related brain differences underlie gender dysphoria. I would feel exceptionally relieved and vindicated to finally have some objective, falsifiable observation to point at that suggests the reality of dysphoria.
Ah, but this is science. How would you feel if it were not vindicated? What if it were not a matter of "born that way"?
Plus, there's a whole ton of confounding business we'll have to look out for. Just off the top of my head, you've got certainty intervals. And wouldn't it be prudent to at least examine the effects of various hormones like testosterone, dihydrotestosterone, estradiol, progesterone on the brain? Say, post age twenty five, pre age twenty five. And so on. Might exist, might not.
We're a long way from being able to have much certainty at all.
I would be shocked if hormones didn't have some effect on the brain considering how much of an effect they have everywhere else. Hell, I know some people who take the "brain sex" angle re. transness speculate that varying levels of hormonal exposure in utero could be what causes gender dysphoria post-utero... in which case, would adult endogenous hormones be responsible for brain structures, or would brain structure dictate which kind of hormones your body expects?
I think, in utero, the effects would be strong and likely visible, but through childhood, puberty, and then young adulthood, I would expect diminishing degrees of impact. Certainly, I've read case studies of various pre-natal exposures to particular chemicals having large and unmistakable impacts on sexual orientation. It's of especial interest to me as one of the very last DES babies.
Brain structure, as in the gross features visible on imaging, would probably not dictate what kinds of hormones you body might expect. I think that sort of thing would occur at a much lower level, right on the receptors of the tissues, and would not be visible.
I think it would be the same as right now, absence of evidence and all that. Would still be short one explanation explaining the observed phenomenon. You're absolutely right that it wouldn't be proof in any sense, just a sign that the avenue would be worth pursuing.
I love that idea! Honestly even if it doesn't show up as a sex-difference doing this same trick might be nonetheless fruitful by asking it to predict say cis woman / trans man.
> I believe this is a clue to where the GP post is coming from: we can't, therefore the study must be flawed.
How so? They are clearly saying they expected this to be possible already, until this article claimed otherwise (by suggesting AI is contributing something new here).
They are saying that this result will not resolve the disputes, because it doesn't address the core thing in dispute: that there are meaningful differences that people actually care about. For example, maybe it is possible to tell the sex based on the shape of the brain. This doesn't mean that men and woman think differently, which is what people are actually arguing about. And the AI in this study can't proof that, because it doesn't give us any further insights into how cognition works.
This point strikes me as actually pretty mundane and obviously correct. The fact that 5 people immediately seem to have misunderstood it (as in, they are not responding to the argument) seems to tell us something about their priors instead.
>They are saying that this result will not resolve the disputes, because it doesn't address the core thing in dispute: that there are meaningful differences that people actually care about.
That's not their claim. Their claim is that there is some flaw ("confounding thing") in the study. They then go on to say that men and women's brains are "indistinguishable." Well, of course you will automatically assume there is a flaw if that is your belief.
There’s some linear and obvious differences like the brain volume. Don’t need ai for that lol. I think I remember that it’s a 20% average difference. Downvote me please
It doesn't, because there is no explanation for how it determines the difference. It's just as likely some confounding thing that's not causally based on the brain, like all these other "AI can tell" studies that don't identify an actual mechanism.
That said I'm surprised that we can't tell men and women apart from the brain. I would have assumed there would be some differences that would make it obvious at least for the average brain. It's more surprising to find they are indistinguishable.