This article mentions the words accuracy/accurate/accurately 10 times, e.g.
"when using his own dating-site-sourced dataset, was accurate at predicting the sexuality of males with 68 per cent accuracy – better than a coin flip"
But accuracy seems like a poor measure for something like this, when the population is highly unbalanced. It's trivial to create a classifier with high accuracy: just outputting 'heterosexual' every time would yield ~90% accuracy on faces of the general population.
This is true, which is why machine learning has long since learned to not even think of what you describe as a meaningful measure of accuracy. If you look at the linked paper [0], you'll find that the author uses the "ROC AUC" metric [1]:
>The ROC AUC score represents the probability that when given one randomly chosen positive instance and one randomly chosen negative instance, the classifier will correctly identify the positive instance
The article didn't mention AUC, so I assumed they were talking about accuracy in the sense people normally mean it, which also matches the definition in the sidebar of the wikipedia link you shared:
People don’t usually look at others completely naked and shaved and devoid of any other things like fashion and style. Why is this not allowed to be used by the AI?
What's really astonishing about this is how opaque the behavior of the CNN is. We have no idea why it is doing what it is doing, and multiple papers and orthogonal investigations are required to figure it out.
"class activation map" is a new term to me, so I looked it up very briefly.
Does it mean simply "what part of the image (or in aggregate, images) carried the strongest signal for categorization"? Or can it also tell you something about the feature set?
"A class activation map for a particular category indicates the discriminative image regions used by the
CNN to identify that category"
So, it's mostly what you said, but the maps you get are specific to the particular trained network you're using; the class activation map tells you, for a given network, for a given image, for a given class, which image regions did that network consider relevant to that class.
This title ('The infamous “AI gaydar” study was repeated – couln't be reproduced') is very incorrect; it does not correctly summarize either the linked article or the actual paper (https://arxiv.org/pdf/1902.10739.pdf) that the article is discussing.
From the abstract on the actual paper:
> Using a new dataset of 20,910 photographs from dating websites, the ability to predict sexual orientation is confirmed... While demonstrating that dating profile images carry rich information about sexual orientation these results leave open the question of how much is determined by facial morphology and how much by differences in grooming, presentation and lifestyle. The advent of new technology that is able to detect sexual orientation in this way may have serious implications for the privacy and safety of gay men and women.
So, the study was repeated, and could be reproduced. However, it's still not clear if there's any merit to the facial morphology hypothesis, vs. the (less worrying) hypothesis that the models are just picking up on differences in grooming and presentation.
13 comments
[ 3.2 ms ] story [ 27.2 ms ] thread"when using his own dating-site-sourced dataset, was accurate at predicting the sexuality of males with 68 per cent accuracy – better than a coin flip"
But accuracy seems like a poor measure for something like this, when the population is highly unbalanced. It's trivial to create a classifier with high accuracy: just outputting 'heterosexual' every time would yield ~90% accuracy on faces of the general population.
>The ROC AUC score represents the probability that when given one randomly chosen positive instance and one randomly chosen negative instance, the classifier will correctly identify the positive instance
[0] https://arxiv.org/pdf/1902.10739.pdf [1] https://en.wikipedia.org/wiki/Receiver_operating_characteris...
The article didn't mention AUC, so I assumed they were talking about accuracy in the sense people normally mean it, which also matches the definition in the sidebar of the wikipedia link you shared:
(TP + TN) / (P + N)
Does it mean simply "what part of the image (or in aggregate, images) carried the strongest signal for categorization"? Or can it also tell you something about the feature set?
Would you suggest and good introductory example to the concept? This was the best I found in my search https://jacobgil.github.io/deeplearning/class-activation-map...
"A class activation map for a particular category indicates the discriminative image regions used by the CNN to identify that category"
So, it's mostly what you said, but the maps you get are specific to the particular trained network you're using; the class activation map tells you, for a given network, for a given image, for a given class, which image regions did that network consider relevant to that class.
For learning more about interpreting neural nets I would highly recommend the posts on distill.pub; https://distill.pub/2018/building-blocks/ and https://distill.pub/2019/activation-atlas/ are probably most pertinent.
From the abstract on the actual paper:
> Using a new dataset of 20,910 photographs from dating websites, the ability to predict sexual orientation is confirmed... While demonstrating that dating profile images carry rich information about sexual orientation these results leave open the question of how much is determined by facial morphology and how much by differences in grooming, presentation and lifestyle. The advent of new technology that is able to detect sexual orientation in this way may have serious implications for the privacy and safety of gay men and women.
So, the study was repeated, and could be reproduced. However, it's still not clear if there's any merit to the facial morphology hypothesis, vs. the (less worrying) hypothesis that the models are just picking up on differences in grooming and presentation.