Facebook Develops Self-Supervised AI System to Learn Without Human Labels
Facebook's AI research division creates SEER, a billion-parameter computer vision model that learns from random, unlabeled images without human curation.
Facebook's AI research division introduces SEER, a billion-parameter computer vision model that applies self-supervised learning to visual tasks for the first time. Unlike traditional systems that require extensive human labeling and curation, this new system learns entirely on its own by analyzing random, unlabeled public Instagram images.
The system utilizes a technique called joint embedding, where a neural network analyzes pairs of nearly identical images. The AI learns to produce similar mathematical vectors for matching images and different vectors for non-matching ones, effectively teaching itself to recognize patterns and objects without being told exactly what it is looking at.
This approach mirrors the recent revolution in natural language processing and aims to give AI systems a foundational sense of common sense. FAIR scientists believe that removing the bottleneck of manual data labeling allows the AI to scale its learning capabilities and adapt to a much wider variety of visual recognition tasks.