The dataset contains images of children’s faces downloaded from websites about autism, which sparked concerns at Springer Nature about consent and reliability.
My point is if someone handed me a tool that could decisively identify autistic kids from photographs I would destroy it if possible and then and advocate for parents never uploading photos of their autistic kids to the internet.
There are people in power who would immediately use this tool to do evil things, to seek to create it makes it clear the creators clearly don’t give a shit about any other goal than “I figured something out!” in their research.
The best current treatments for autism depend on knowing as early as possible that a child needs specialized help to have the best outcomes. If there was a tool that decisively identified autistic kids and you destroyed it I would rank you as a particularly heinous monster.
This tool wouldn’t work very well either. Ignoring the problems of collecting and verifying the validity of the data (and those are pretty big, serious problems), 2,900 images isn’t really enough to train an accurate image classifier. Especially not one that I would be comfortable using in a medical context.
Using the tool would probably result in a ton of false positives, and I’d bet the model would be overfitted to its data. Of course, I don’t think that would prevent people from using something like that nefariously.
My point is if someone handed me a tool that could decisively identify autistic kids from photographs I would destroy it if possible and then and advocate for parents never uploading photos of their autistic kids to the internet.
There are people in power who would immediately use this tool to do evil things, to seek to create it makes it clear the creators clearly don’t give a shit about any other goal than “I figured something out!” in their research.
The best current treatments for autism depend on knowing as early as possible that a child needs specialized help to have the best outcomes. If there was a tool that decisively identified autistic kids and you destroyed it I would rank you as a particularly heinous monster.
This tool wouldn’t work very well either. Ignoring the problems of collecting and verifying the validity of the data (and those are pretty big, serious problems), 2,900 images isn’t really enough to train an accurate image classifier. Especially not one that I would be comfortable using in a medical context.
Using the tool would probably result in a ton of false positives, and I’d bet the model would be overfitted to its data. Of course, I don’t think that would prevent people from using something like that nefariously.