Facebook Unveils SEER: AI Model That Learns to Recognize Images Without Labels
Facebook introduces SEER, a self-supervised AI model that teaches itself to identify objects in photos using unlabeled Instagram data. The breakthrough system outperforms existing models while significantly reducing training time.
Researchers at Facebook create a groundbreaking image recognition model called SEER that teaches itself to understand photos without relying on carefully labeled datasets. The system, whose name stands for "Self-SupeERvised," processes one billion public, unlabeled images from Instagram and learns independently as it works through the massive data collection. Facebook considers this a major breakthrough that potentially sparks a revolution in the field of computer vision.
Self-supervised learning offers significant advantages over traditional AI training methods because it eliminates the painstaking, time-consuming need for humans to label data manually. By skipping this step, developers train these models much faster and utilize much larger, more diverse datasets. SEER demonstrates these benefits clearly by leveraging an algorithm called SwAV, which rapidly groups images with similar visual concepts and reduces required training time by six times compared to previous state-of-the-art systems.
The SEER model incorporates a highly scalable architecture known as RegNets and an open-source library called VISSL to achieve impressive results. When tested on the industry-standard ImageNet object recognition benchmark, SEER achieves an 84.2% classification accuracy score, outperforming existing AI models. Facebook states that this achievement highlights the future of artificial intelligence, where systems learn directly from the information given to them rather than relying on human-annotated data.