Facebook Builds Massive AI by Training on One Billion Instagram Photos
Facebook develops SEER, a billion-parameter computer vision system that learns to recognize objects by pretraining itself on a billion unlabeled Instagram images before fine-tuning on labeled data.
Facebook trains its most advanced semi-supervised computer vision system, known as SEER, on a massive dataset of one billion public images sourced from Instagram. The convolutional neural network contains over a billion parameters and successfully identifies various objects in pictures, such as bicycles, bananas, and umbrellas, by converting visual data into descriptive text.
Unlike traditional models that rely entirely on human-labeled data, SEER partly teaches itself through a self-supervised approach. The system first uses an algorithm called SwAV to group similar Instagram photos together without any human intervention, and then developers fine-tune the model using a million labeled examples from the ImageNet dataset to teach it how to properly caption the grouped images.
This semi-supervised method proves significantly more efficient than manually labeling a billion images, saving developers countless hours of data preparation. By combining unsupervised pretraining with supervised fine-tuning, Facebook follows a growing machine learning trend that mirrors the training process of other major AI systems like OpenAI's GPT-3.