A couple of questions about Jeremy's post on the homosexual recognition study

Last night I read Jeremy’s post on one ML study: http://www.fast.ai/2017/09/13/kosinski/

I totally agree with the content of this post, just have a couple of questions:

  1. Specifically, each image was turned into 4096 numbers, each of which had been trained by University of Oxford researchers to be as good as possible for recognizing humans from their faces.” What are these 4096 numbers?

  2. "__They compressed those 4096 numbers down 500 using a simple statistical technique called SVD._… " Does this step improve the model accuracy? Or it only reduce the computation resources requirement?

  3. …and they then used a simple regression model to map these 500 numbers to the label (gay or not).” I think the CNN models can be directly used for image classification, right? Why do they bother to use another regression model?


The study in question uses VGGFace, which is based off of VGG. Presumably they’re using the final layer before the softmax output as a feature vector. I’m not sure why they would compress that with SVD and use regression. You’re correct in your comment that CNNs can do this and I would have thought transfer learning would be the way to go here, although I do share Jeremy and Rachel’s concerns on the study as a whole.


1 Like

Thank you!