Facebook self-supervised computer vision model promises object recognition breakthrough
Facebook has announced the development of a new computer vision model, named SEER (SElf-supERvised). SEER has been pre-trained on a billion public (non-EU) Instagram images, and is able to make inferences between the data’s parts unlike most CV models who learn from pre-labelled datasets, reports Venture Beat.
“Self-supervised learning has incredible ramifications for the future of computer vision, just as it does in other research fields. Eliminating the need for human annotations and metadata enables the computer vision community to work with larger and more diverse datasets, learn from random public images, and potentially mitigate some of the biases that come into play with data curation. Self-supervised learning can also help specialize models in domains where we have limited images or metadata, like medical imaging. And with no labor required up front for labeling, models can be created and deployed quicker, enabling faster and more accurate responses to rapidly evolving situations,” Facebook wrote in a blog post.
Self-supervision is believed to be key in the step away from machine learning, and towards human level intelligence. It could improve speech and object recognition, among other AI applications. A range of issues related to dataset collection and annotation have also plagued biometrics development, particularly in facial recognition.
Instagram’s terms of service allow the company to use data uploaded to it in almost any way, but as OneZero notes, the avoidance of images from European users is likely an attempt to avoid falling afoul of GDPR.
Images do not incorporate semantic concepts, like words do, therefore designing a model that is able to make these inferences required Facebook researchers to use a convolutional network (ConvNet) that was big enough to learn every visual concept from the images. Because the dataset does not use labeling, Facebook plans to automatically populate it with new images every 90 days.
Development of SEER included the use of several components of architecture; an ultra-fast algorithm called SwAV, as well as RegNets (ConvNet); capable of scaling billions of parameters without compromising run-time or accuracy. According to Facebook, the model outperformed the most advanced state-of-the-art self-supervised systems.
Facebook software engineer Priya Goyal says that use of individuals’ Instagram pictures for research is stated in Instagram’s data policy, therefore there was no chance for people to opt-out of this data use. However, Goyal mentions that Facebook does not plan to share the images or the SEER model itself due to potential biases, further described in the full research paper.
Facebook has been implicated in multiple lawsuits regarding collection and use of individual’s biometric data, and allegedly failing to notify users of collection of data and receive explicit consent.