Why is it that we only find certain types of people attractive? It harks back to our college days when we were attracted to a specific individual only to be rebuffed and ticked off by our friends as a “poor” choice. Teenage insensitivity and envy aside, why is it perfectly normal to find someone attractive and yet have people close to us, decry our judgment? More importantly, perhaps, is there a definitive pattern to this? 

Apparently, there is and it is influenced by cultural and psychological factors. Notwithstanding the subjectivity, researchers have successfully gotten AI to understand these kinds of patterns and these systems can now create portraits on their own – tailored, based on individual “taste” which is really about brain responses that can be recorded via electroencephalography (EEG).  

GANs or Generative Adversarial Networks is an approach to generative modeling, using deep learning methods, such as convolutional neural networks. Generative modeling is unsupervised learning that discovers and learns patterns from input data. Subsequently, it can generate outputs as if they were drawn from the input data set. Under generative models, two sub-models work together leading to a zero-sum game. The generator model (under supervised learning) is trained to generate new examples. And, the discriminator model tries to classify the examples generated as fake or from the domain. This goes on till the discriminator model is fooled into believing that half of the time, the cases generated are plausible examples. The two models (generator & discriminator) are trained together in a zero-sum game – hence adversarial.    

GANs is a rapidly changing field that delivers on the promise of generative models. It works across a wide range of problem domains and most notably in image-to-image translation tasks. 

Now let us go back to our study.

Initially, the researchers used GANs to create hundreds of artificial portraits. Thirty volunteers were shown these images and were instructed to pay great attention to faces that were found attractive. Like in Tinder, the dating app, they swiped right (or is it left, whatever?) when a face was found to be attractive. Simultaneously, the brain’s responses were mapped. 

Whew, I never knew dating was so complex (and intrusive)!

The researchers at the University of Helsinki and the University of Copenhagen (where the experiment was conducted) were able to interpret users’ opinions on the attractive quotient of each image. We must bear in mind that it’s easier, rather much easier to teach a machine to identify a blonde or a brunette than classifying whether the image is attractive or not – because this is in the deep realms of subjectivity. Interpreting these “views” led to the generation of new images for each participant and predicting what they would find attractive. In a double-blind, the volunteers communicated their preferences with what was shown (created through the model) as attractive or not – and voila, the accuracy was a high 80%. 

One may well ask at this stage – what kind of practical applications are we likely to see? Well, it has the potential of identifying unconscious attitudes (biases even) especially amongst people who are decision-makers. This will take us a step closer to responsibly using AI.  

Next time you find someone attractive just remember there’s nothing to get excited about – it’s just your brain waves acting funny! And, of course, those Helsinki guys will be on your case in a jiffy. It’s they who deserve all the excitement.