ImageNet Roulette: Find Out What a Neural Network 'Thinks' of You

Can a neural network recognize faces? Yes, of course. With enough training, it might not miss a single detail, reaching chilling levels of accuracy. However, when we ask that same neural network to share its opinion about what it sees… the story changes. ImageNet Roulette is a project based on a neural network trained using the 'Personas' dataset from the ImageNet database, and with the help of 2,500 additional labels, it tries to classify people.

Faces, Neural Networks, and Tags

For better or worse, one of the first things we've done with neural networks is process people's faces. From tags on Facebook to real-time recognition used by Chinese authorities, there seems to be no limit. We even let artificial intelligences invent faces almost without supervision, in a kind of digital fusion out of control.

Can AI Put a Label on You?

So… these networks can identify and create faces, but can they classify them? I'm not talking about superficial details like size and color, but more 'human' characteristics, so to speak. There's always someone with a face of 'doctor', 'TV host', 'policeman', and other labels we don't hesitate to place on others. Could a neural network do the same? ImageNet Roulette explores the answer to that question.

The creators of ImageNet Roulette presented this project as 'a provocation' designed to analyze the way a human being is classified by artificial learning platforms. The training source for this 'roulette' is the 'People' dataset in the ImageNet database. To it are added a total of 2,500 labels derived from WordNet, and the rest is letting the network draw its conclusions.

How ImageNet Roulette Works

Generally, ImageNet is a dataset applied to object recognition. However, as a preview of the 'Excavating AI' research project by Trevor Paglen and Kate Crawford, they decided to limit the training to the People section of ImageNet, with its 2,833 subcategories included. ImageNet Roulette adopts an open Caffe deep learning framework. When the user uploads an image, it first tries to detect the presence of a face, and if it finds one, it is transferred to Caffe for classification. The last phase is based on returning the original image, accompanied by the assigned labels.

The creators warn that the labels can be offensive or misogynistic, since that content is part of the WordNet base. The good news is that ImageNet Roulette does not store any image, so you can upload a photo of yourself and see its interpretation without risks. Give it a try!

Official site: Click here