Experts from the Beijing Academy of Artificial Intelligence announced the creation of a generative deep learning neural network Wu Dao 2.0 with record properties. Its power is much higher than that of its closest competitors – Open AI GPT-3 and Google Switch Transformers.
The advantage of the new neural network is power. They are measured in the number of training parameters: that is, this is the number of factors that the neural network can use in the course of its work.
The Open AI GPT-3 neural network uses 175 billion parameters – this power allows it to generate meaningful texts, create music and program code.
Wu Dao 2.0 has more than 1.75 trillion parameters – 10 times more powerful than GPT-3.
A new development can perform different functions, it:
- writes an essay,
- poems and couplets in traditional Chinese,
- recognizes images and generates them by verbal description,
- imitates speech,
- creates culinary recipes,
- predicts the three-dimensional structure of proteins.
During the training of the neural network, the developers used almost 5 TB of images and texts, including 1.2 TB of texts in Chinese and English.