The first algorithm that can learn things independently would have been developed. The honor would go to Meta.
Researchers at Meta say they have developed an algorithm that can train itself with multiple sources of information. Text, images and speech fragments can all be used by the algorithm as learning material. This would open the way for more advanced AI that can perform multiple different tasks.
Combination previously not possible
According to the researchers, the developed algorithm is unique because it can use anything from the environment to train itself. Previously, it was not possible to use both text, sound and images for one algorithm and a choice had to be made. With the newly developed algorithm, data2vec, that is a thing of the past, according to the researchers.
Moreover, data2vec always required providing mountains of labeled input data. That could then be labeled text or sound fragments, which taught the algorithm to focus on certain things. After that learning process, the algorithm could predict words for text fragments or predict sounds for sound fragments.
Data2vec
At data2vec, Meta researchers have developed a new method. Instead of having the AI model focus on words, sounds or images, the algorithm lets AI models predict their own representations of the input data. Humans likewise estimate their environment by combining different stimuli. For example, we only experience a night scene in a movie as exciting if the music makes that clear to us.
Mark Zuckerberg, founder of Meta, raves on Facebook about data2vec: “Meta AI research has built a system that learns from speech, sight and text without the need for labeled training data.” According to him, in the future, the algorithm can be perfectly built into AR glasses to control the AI assistant.
The code and several already trained models Meta researchers are now sharing with other researchers in the field. In this way, data2vec can serve as a starting point and be further refined by researchers.
read also