Facebook trained its SEER AI and now it is able to recognize images and objects by itself.
This AI was trained with more than a billion public images extracted from Instagram. As commented by Facebook, SEER has managed to outperform existing AI models in an object recognition test with a score of 84.2%.
Currently, Instagram is used as a public photo album by many people. Therefore, with that database, Facebook has access to a huge and varied catalog of options, including people of a wide range of ages and body features, animals, landscapes, food, and a long etcetera. A significant sample of this information was made available to the AI for training.
Self-supervised computer vision is used on SEER
An AI that operates autonomously acquires after training the ability to establish relationships between objects or variables. In this case, since we are dealing with images, SEER’s task is to manage to label them appropriately without having any context data.
This marks an important difference in relation to other image recognition systems that, although they also work thanks to an AI engine, they follow a supervised dynamic, depending on previous labels or classifications to replicate them.
This AI makes use of two key elements: one is an algorithm capable of learning data from a large number of random images, without metadata or annotations; while the other is a convolutional network large enough to capture and learn all the visual concepts from these large volumes of complex data.
This contribution could enhance the automatic generation of image descriptions and alternative texts for visually impaired people, filtering inappropriate content, and even automatically categorizing the items listed in Marketplace in a better way.
Facebook’s AI is only a prototype for now
At the moment, SEER is only a prototype. To boost it, Facebook announced that it will release part of the code of this software under an open-source license, generating an instance of experimentation for developers and researchers, strengthening the operation of this new image recognition technology.
At the beginning of the SEER presentation, Facebook emphasized that “the future of AI is in creating systems that can learn directly from whatever information they’re given — whether it’s text, images, or another type of data — without relying on carefully curated and labeled data sets to teach them how to recognize objects in a photo, interpret a block of text, or perform any of the countless other tasks that we ask it to.”
The official announcement was released by Facebook through its artificial intelligence projects blog, where more technical aspects about the presented AI system are shared.