A Fascinating Machine Learning Approach to Generating Faces in Advertisements
- Researchers from the University of Pittsburgh have developed a model that generates faces for advertisements
- Their approach involved first analyzing what makes ads persuasive and then using their GANs technique accordingly
- The results are promising, and were also extended to generated objects other than faces
Advertisements are a powerful tool that affect our habits and decisions. For years, we have been watching them everyday on television, and splattered across posters and print media. These ads have become even more pervasive in the digital age.
What makes these ads different is the way they are portrayed. Regardless of what the ad is about, we see all kinds of faces attached to it – from beauty products to beverages, there is usually a face next to the actual product. So instead of splurging out money, what if you could use machine learning to generate faces according to the advertisement requirements?
That’s exactly what a couple of researchers from the Computer Science department of the University of Pittsburgh went about doing in their research. They took a multi-step approach to understanding what makes advertisements persuasive and then generating faces which appear as if they are coming from different types of ads.
The two researchers built a conditional variational autoencoder for this purpose. This techniques makes use of predicted semantic attributes and facial expressions as the supervisory signals during the training process. As you can see in the above and below images, the model was able to generate unique facial images according to the category of the advertisement. In the below image, the models you see on the x-axis are the ones trained and/or modified by the researchers.
This model outperformed several baseline models, including two previous state-of-the-art GANs built for transforming faces. Interestingly, the researchers extended their study to generate objects as well. The results were pretty impressive in terms of how detailed each image looked.
You can also read about this study in more detail, including the mathematical notations, in the recently published research paper.
Our take on this
The first question that came to my mind was – will this replace humans in the long run? It’s too early to say but this research certainly encourages that line of thinking. It gives business, especially those geared towards marketing and advertising, food for thought.
GANs have been around for a few years but are rapidly gaining popularity these days. It’s intriguing to see how researchers are using ML to first analyze what makes ads persuasive, and then leverage deep learning algorithms to take advantage of the said analysis. It gives you a window into the thinking of top experts in this field.
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!