Digital Reasoning is a company that works in cognitive computing and it has announced that it has smashed a record set by Google for neural networks. Digital Reasoning has trained a neural network with 160 billion parameters. Google’s record was for training a neural network with 11.2 billion parameters. The network with 160 billion parameters also surpasses Lawrence Livermore National Laboratory’s record of 15 billion parameters.
Digital Reasoning will publish details on the research that goes along with the neural network in the Journal of Machine Learning and Arxiv. Details on the record setting neural network will also be shared at the 32nd International Conference on Machine Learning in Lille, France on July 6-11.
A neural network is a computer system that is modeled on the human brain and has the ability to gather raw data, process that data, and react to the data it has processed. In the project, the team of researchers evaluated their network using neural word embeddings on word analogy accuracy. Neural networks have the ability to generate a vector of numbers for each word in its vocabulary.
That fact lets the researchers perform something known as “word math” using the network. One example given is giving the network input such as “king minus man plus woman” with the neural network coming up with the answer “queen.” Digital Reasoning was able to achieve an accuracy rating of 85.8% with its network compared to Google’s best accuracy of 76.2%.
SOURCE: Digital Reasoning