Facebook has open-sourced its Python library AugLy, which aims to help AI and machine learning researchers use data augmentations to assess and improve their AI/ML models. AugLy provides data augmentation tools to create samples to train and improve different AI systems.
AugLy’s library combines image, audio, text, and video. In recent years, it’s become popular in several types of AI research fields.
Augly provides over 100 data augmentations based on actual images, videos, and other media on platforms like Facebook, Pinterest, and Instagram.
Data sets and models are becoming more multimodal. Accordingly, having a project’s data under one library and API can be extremely helpful.
On top of that, combining different modalities with real-world augmentations can help AI models better understand complex content.
Data augmentations provided by AugLy are derived directly from data transformations on Facebook.
As a result, AugLy will be helpful for those working on models or data related to social media.
AugLy has four different sub-libraries. Each are associated with a different modality.
However, each library has a similar interface. Accordingly, AugLy can generate metadata understand how their data is being transformed. Facebook has data augmentations from different existing libraries.
Data augmentations are vital to enhance the quality of AI models. For instance, if models need to be able to ignore unimportant data, augmentations will help them learn to focus on the critical aspects of a dataset for a particular use case.
Facebook asserts that an important application detects duplicate content copries. For instance, misinformation can repeatedly appear in different forms of audio, text, etc.
After AI models are augmented with AugLy inputs, they can help detect false information. This will, in turn, help in proactively moderating activity to prevent the uploading of infringing content.
AugLy’s library will not only train models, but also help determine the durability of models pertaining to a set of augmentations.
AugLy has been used during the Deepfake Detection Challenge to determine the robustness of Deepfake detection models.
Much infringing content online is done by those who change it as a way to skirt around automatic detection systems.
This means AugLy can help AI researchers on a variety of tasks, which can include object detection models, voice recognition, hate speech, and other applications.