Facebook has published the code for their self-supervised learning algorithm and made it publicly available for use. The company is also making the components of the VISSL library and the SwAV algorithm available for public use. However, it is not making the Instagram data set or the trained SEER A.I. model available. This is disappointing, as it could help many people use the new technology to improve their own Instagram photo captioning.
Self-supervised learning:
A recent study by the Facebook AI team revealed that its new SEER A.I. model has outperformed state-of-the-art self-supervised learning models in recognizing images. The new method was trained on a billion public Instagram photos and outperformed existing self-supervised learning models by more than one percentage point. Those findings are good news for data scientists, but it’s not clear whether the research will be useful in the real world.
SwAV algorithm:
Developed by Facebook AI, the SwAV algorithm for its SEER Instagram model focuses on clustering similar visual concepts. The new algorithm has 6x less training time than previous methods, enabling it to be trained in half the time. The new algorithm also reduces memory requirement per graphical programming unit, which speeds up the training process of any model.
VISSL library:
The Facebook AI team is encouraging machine learning developers to share their knowledge and ideas. For this reason, they’ve published an open source library called VISSL, which is used to develop SEER. The library aims to reduce the memory requirements of any model and increase its training speed. Here are some of the techniques they used to make SEER more powerful:
Scale of dataset:
This new study uses one billion publicly available Instagram images to train an AI system that can detect objects without labeling or annotation. The system achieved top object detection accuracy and was trained with 512 NVIDIA V100 Tensor Core GPUs and 32GB of RAM for 30 days. The researchers used mixed precision and gradient checkpointing tools from the PyTorch library to reduce memory consumption and increase training speed.
Results of test:
The results of the latest AI challenge from Facebook’s AI team are promising: the new SEER system outperforms the best self-supervised systems on ImageNet’s object recognition test. The benchmark test requires the model to identify objects from images in a database of thousands of photos. This new algorithm utilizes an older algorithm called SwAV, which leverages similarities in images to group them together. It has improved significantly over the previous state-of-the-art self-supervised systems, with 6x less training time.