Google is adding a new feature to Allo that harnesses the power of neural networking to give you full freedom of expression by turning your face into a pack of stickers that you can paste into any conversation. The feature draws on many of the same conventions as Google’s Deep Dream technology to figure out the closest approximation possible for a user’s looks, using a wealth of artist-created resources to construct cartoon representations of users. The pool or cartoon-style resources used to represent users is drawn through a community-driven process that ends with an artist picking from a few different pictures of users with a common feature, then drawing the best possible mutual representation of that feature. Google has not announced exactly when this feature will actually be available in the app.
This breakthrough in machine learning was made possible by Google’s engineers working to teach the computers how to see users in the context of the things around them, taking into account the qualitative features that may introduce subtle changes, and finding the universal truth among them. Using large neural networks built with no specific purpose in mind, Google ran the trials and eventually found that certain nodes had a knack for noticing small details outside of their assigned purview. Google built on that to teach all of the nodes how to see the bigger picture, so to speak. This enabled the neural network to not only see all the details of a given user selfie for what they were, but to objectively compare those details and come up with a ground truth from them, which is what makes the cartoon conversion process possible.
In order to avoid stepping into the uncanny valley, Google took a twofold approach. First, they kept the depictions strictly non-detailed and cartoonish, ensuring that users could essentially project themselves onto an avatar that looks vaguely like them. Second, the feature will allow users to customize an avatar before it’s put together and turned into a sticker pack. This allows users to settle on an appearance that they’re personally comfortable with. The wide range of features on offer can create literally quadrillions of different unique appearances, so the result is something that’s just unique enough to be a good representation of a user, while being generic and stylized enough to not stumble into the uncanny valley. Biases in the machine, a rapidly rising social issue intrinsically linked to the “humanization” of AI, are fought by continuous training and by using resources that are androgynous and generic enough to represent a user in a more general sense and letting users pick the final result on their own.
