Understanding the Science of Beauty

There is an undeniably strong bias towards Caucasian faces in just about any currently existing Machine Learning (ML) algorithm, from CNN classifiers to GAN generators. A recent study by Karkkainen and colleague found that Hispanic and Middle Eastern ethnicities were greatly underrepresented in most existing ML datasets,

with Black and South Asian ethnicities being not too far behind. This bias in ML is common knowledge and has heavily stigmatized the large majority of AI research to the general public, especially with anything that claims to look at beauty as objectively as possible.

Based on our prototype facial assessment tool  which uses a similar CNN (convolutional neural network) to classify 17 different cosmetic conditions, this version 2 will feature improved accuracy with much better diagnostic feedback and transparency. Our version 2 will use a newer image segmentation model to split key regions of the face apart

to be assessed individually to healthy examples, providing much more accurate results. Version 1 used a holistic evaluation which looked at the face as a whole. As always, it will feature an ethnically diverse training protocol.

Using the current research literature on face shape detection, we’ve devised a novel approach to recommending the right hairstyle for faces based on face shape, hair and skin texture and androgenicity. In doing so, users can not only try on different

hairstyles using a StyleGAN approach of morphing a target face with the desired hair with the users, but they will also be provided with the correct aesthetic recommendations so that the process of finding a suitable hairstyle is less based on trial and error.