DEV Community

Detecting academics' major from facial images

Ferdinand Mütsch on January 02, 2019

The Idea A few months ago I read a paper with the title "Deep Neural Networks Are More Accurate Than Humans at Detecting Sexual Orientat...
Collapse
 
ben profile image
Ben Halpern • Edited

First of all, this is not a scientific study, but rather a small hobby project of mine. Also, it does not have a lot of real-world importance, since one might rarely want to classify students into four categories.

I really like that this is sort of a humble project without a big end game. That’s where much of the most interesting findings come from. It's also where the most disastrous outcomes happen. In an attempt to reduce bias or find "factual results", there's a lot that can go wrong here. So I like the self-awareness here.

Collapse
 
elmuerte profile image
Michiel Hendriks

I would qualify these kinds of experiments as ethically dangerous. This is basically a modern form of phrenology. There are already companies in the while which using machine learning to produce "reports" about people based of pictures or videos. (For example companies which facilitate job interviews via video calls.) These models usually end up being nothing short of badly discriminatory.

If for example you would train a machine based on the pictures of the creators of programming languages and the success of said languages you will probably get the conclusion: successful languages need creators with facial hair. Which would result in rejection of languages created by people without a beard or mustache, like women.

As I commented in What is your personal Programming ethics?
This question should always be in your head when creating things:

What would the worst people do if they got hold of this?

Collapse
 
yorodm profile image
Yoandy Rodriguez Martinez

Agreed, see discussion here

Collapse
 
antimon3 profile image
Flo Roform • Edited

Hi Ferdinand, very nice project, good work!

It would be great to know the classification accuracy on pictures that you yourself take on a campus tomorrow. I can imagine that your classifier did not only learn the features of a face but also certain camera properties (contrast, brightness, …) as you were crawling websites of institutes that tend to take pictures of their staff with the same camera. What do you think?

Anyway, great project!

Collapse
 
n1try profile image
Ferdinand Mütsch

Yes, you're absolutely right. Also, some institutes might take photos of their staff in front of the same background, with same lighting, etc. These are features that might bias the model.
Taking pictures on campus by myself is infeasible, but I could try to have test data, which consists purely of pictures from "unseen" institutes / departments.

Collapse
 
ferricoxide profile image
Thomas H Jones II

Might have had better results with full-body images (assuming you could scrape nearly as many images). It might just be a US thing, but clothing-choices are usually a lot more indicative of tastes which, in turn, can influence areas of academic interest.

Collapse
 
n1try profile image
Ferdinand Mütsch

Yes, I agree! With full-body images, a model might even learn things like posture. There would be a lot more information to extract knowledge from. But probably you would need a different approach for gathering enough training data.

Collapse
 
kellydimarco profile image
kellydimarco

This is true . It come after lots of research work. At oxbridgeessayshelp.co.uk/ we are here to provide you all solutions for your Academics Research Writings, Reports, Assignment,Article, Thesis and Dissertation in a professional manner to mark yourself on the top in your career as you deserve the best. Our expert writers will deliver you the best according to your needs and budget.

Collapse
 
stereobooster profile image
stereobooster
Collapse
 
gokatz profile image
Gokul Kathirvel

I loved this blog without even fully read! Just an awesome idea 🥳🥳

Collapse
 
epogrebnyak profile image
Evgeny Pogrebnyak

In (1) false negatives, (2) false positives and (3) true positives - top row is all men, bottom is all women. Is that coinidence? Maybe women pay more attention to thier appearance on the photo?

Collapse
 
n1try profile image
Ferdinand Mütsch

I would say one interpretation of that fact is that the model is pretty confident about female economists, but unsure about male ones. However, I would have to look deeper into that.