Learning and updating constantly is what has helped the human race come from being cave dwellers to masters of technology. To progress to a new frontier, researchers believe that Artificial Intelligence (AI) too should have cognitive functions that are as quick and accurate as us. However, AI so far has needed a lot of guidance and handholding as it requires to be trained on labelled datasets. This is where self-learning AI has made headway. 

Facebook researchers announced in mid-March that they are working on an AI project called ‘Learning from Videos’, that is capable of self-learning from publicly available videos. It uses the audio, textual and visual data to add to its growing understanding. The project focuses on giving content recommendations for users, ensuring content policy enforcement and enhancing AI's ability to learn like humans. So far, the researchers claim that the project can learn and recognize any video content. 

The 'Learn from Videos' developers have trained the project to absorb a vast array of data from different regions and cultures across the world. This, the developers assume, will help create a platform that can predict user habits while on Facebook products. The developers have said that the project has adopted a very transparent privacy policy that ensures that the project undertakes steps that are ethical. The developers have further confirmed that the content creators are sure about using this feature since it focuses on user trust and safety guidelines.

The project has already tested two of its products - semi-supervised and self-supervised AI. Developers say that this feature can also notify and decrease instances of hate-speech. This product's first experiment was first recognised through Instagram Reels' recommendation system. 

The Reels are the proof of concept that the self-supervised learning AI can examine common themes that are trending on the applications and give suggestions about the trending videos related to the overall content. " To achieve this, we leveraged Generalized Data Transformations (GDT), our state-of-the-art method for building video embeddings, which systematically learns the relationships between the sound and images in a video. Since building this technology last year, we’ve pioneered the large-scale application of GDT to the representation of Reels data, by training a series of models on a data set of millions of Reels and videos from Instagram," says the official Facebook blog.

The developers have used the latest technique for learning speech representations, called wav2vec 2.0, which masks a portion of the speech and learns to predict the remaining masked units. "To provide an idea of the speed of progress, wav2vec 2.0 and self-training requires only 10 minutes of transcribed audio to achieve very good speech recognition results on the LibriSpeech industry benchmark. The same results required nearly 1,000 hours of transcribed audio just one year ago," elaborates the blog. The model was applied on millions of unlabelled videos and only 100 hours of labelled data. The self-supervised learning products have shown 20% reduction in speech recognition errors which could aid in providing more accurate auto-captioning. 

The developers said that the AI is attempting to filter out material that has already been viewed by the users, enabling users to find newer content on Facebook platforms while the AI sharpens its abilities to learn to tell the difference between similar content. The developers will further develop the wav2vec 2.0 to include more data and language. "As part of these efforts, we’re currently working on training a multilingual model with millions of hours of speech from 25 languages," declared the writers of the blog. 

As per the team, the next major challenge is programming the project to draw memorized audio and visual input and then equate the two according to a general theme.