Nvidia announced its breakthrough in artificial intelligence in language understanding which could enable real-time conversational AI in a variety of software.
The real-time conversational AI is essential for companies that want to develop chatbots as well as virtual assistants that can have discussions with actual persons while showing “human-level” understanding, stated Nvidia.[wpinsertshortcodead id=”bzyqm5d3e04029f48f”]
In a press briefing, Bryan Catanzaro vice president of applied deep learning research at Nvidia said:
Conversational AI has tons of applications all over the world. But it poses a lot of challenges. The industry has been moving toward much larger language models, but they’re difficult to train and difficult to deploy.
Nvidia’s recent milestone includes breaking the time in training BERT (Bidirectional Encoder Representations from Transformers) from several days to just 53 minutes. BERT is one of the world’s most innovative AI language models, and a state-of-the-art model widely regarded as a good standard for natural language processing. Nvidia was also able to cut the time to complete AI inference that is just over two milliseconds. It is sufficient to handle any sort of fast-paced discussion that a human would expect.
It set a world record for training BERT-Base in under one hour, which usually takes weeks, by utilizing optimized software and its DGX SuperPOD system, Nvidia stated. Moreover, Bert Inference with just two milliseconds latency by Nvidia’s Tensor RT platform, which is within the 10-millisecond limit needed for human-level accuracy.
As per Catanzaro, with the recent advancements, Nvidia is looking forward to power the next wave of conversation AI and the company currently made some strong development.[wpinsertshortcodead id=”zxikm5d3e04a8f1451″]
To achieve the breakthroughs in conversational AI, Nvidia made many optimizations to its AI platform, which are now available to developers. The optimizations are made open-source on GitHub which includes the new BERT training code with PyTorch and TensorRT optimized BERT sample.