While artificial intelligence (AI) is often inspired in; the image of a human, so far, AI has been able to emulate only one skill at a time - different algorithms can learn to identify images, produce speech and respond to human facial expressions, but not as generalists like humans. However, researchers at Meta AI, formerly known as Facebook AI Research, have created a single learning algorithm called data2vec that can train neural networks, the first high-performance self-supervised algorithm that works for multiple modalities.

Traditionally, to train AI models to perform a task, they are first fed with humongous datasets of labelled information. However, these databases are hard to create - costing significant money as well as time - especially to train next-gen AIs. Currently, self-supervised models look promising as they can 'learn' from large quantities of unlabeled data, like books or videos of people interacting, and build their own structured understanding of what the rules are of the system, very much like how humans gather knowledge and build skills - intuitively.

Meta AI's data2vec, like other self-supervising learning algorithms, enables neural networks to identify patterns in the data sets without being guided by examples. GPT-3, one of the most successful natural language processing AI uses a similar technique where it trains on vast bodies of unlabeled text scraped from the internet.

Data2vec uses two neural networks - the teacher and the student. "Our method uses a teacher network to first compute target representations from an image, a piece of text, or a speech utterance. Next, we mask part of the input and repeat the process with a student network, which then predicts the latent representations of the teacher. The student model has to predict representations of the full input data even though it has a view of only some of the information. The teacher network is identical to the student model but with weights that are slightly out of date," states the official Meta AI announcement.

The approach was tested on ImageNet computer vision benchmark for a vision where it performed than existing methods, and it outperformed Meta AI's previous self-supervised algorithms wav2vec 2.0 or HuBERT on speech. It also outperformed Meta's RoBERTa on the popular GLUE benchmark suite, to read text.

Mark Zuckerberg is already dreaming up potential metaverse applications. "This will all eventually get built into AR glasses with an AI assistant," he posted to Facebook today. "It could help you cook dinner, noticing if you miss an ingredient, prompting you to turn down the heat, or more complex tasks."

The applications for a general AI are plenty, and Mark Zuckerberg is already envisioning products with data2vec. "This will all eventually get built into AR glasses with an AI assistant," he posted to Facebook today as per reports by MIT Technology Review. "It could help you cook dinner, noticing if you miss an ingredient, prompting you to turn down the heat, or more complex tasks."

“The core idea of this approach is to learn more generally: AI should be able to learn to do many different tasks, including those that are entirely unfamiliar,” wrote the team in a blog post. “We also hope data2vec will bring us closer to a world where computers need very little labelled data in order to accomplish tasks.” The code for data2vec is open-source; it and some pretrained models are available here.