Microsoft has revealed its largest deep learning language model, the Turing Natural Language Generation (T-NLG), which is claimed to have a record-breaking 17 billion parameters.
The T-NLG, according to Microsoft, outperforms the largest deep learning models to date: the University of Washington’s Grover-Mega and Nvidia’s MegatronLM, which possess 1.5 and 8.3 billion parameters, respectively.
According to Microsoft, the T-NLG is capable of completing unfinished sentences, as well as generating direct answers to questions and can create summaries of documents fed into it.
Microsoft also claims that the model has the ability to directly answer the question with a complete sentence.
“This capability is more important outside of web search—for example, this can power AI assistants to intelligently respond when a user asks a question about their personal data such as emails or Word documents,” explained Microsoft’s applied scientist Corby Rosset.
He also thanked the DeepSpeed Library and the ZeRO optimiser for producing “breakthroughs” without which “this work would not be possible”.
Providing an example of how the T-NLG works, the language model introduced itself by generating a summary of its skills:
“Turing Natural Language Generation (T-NLG) is a 17 billion parameter language model by Microsoft that outperforms the state of the art on many downstream NLP tasks. We present a demo of the model, including its freeform generation, question answering, and summarization capabilities, to academics for feedback and research purposes,” said the T-NLG.
