Maluuba, a Canadian artificial intelligence startup, created a system that can read, comprehend and reason, almost as well as humans.
Their deep learning-based program called EpiReader is designed to solve a specific kind of comprehension task: a word is removed from a block of text and then the system is able to determine the missing word based on the context.
Trained on TITAN X GPUs with the cuDNN-accelerated Theano deep learning framework, the EpiReader was tested on two large collections of texts — CNN / Daily Mail collection (over 300,000 articles) and Children’s Book Test (98 classic children’s books). The system scored 74% and 67.4% accuracy on the two datasets which experts say is the highest results seen yet for these two tests.
Watch EpiReader in action.
“We really wanted to incorporate the higher level reasoning function of human beings into EpiReader, the way we go about thinking about the world,” Adam Trischler, Maluuba’s head of research. Mohamed Musbah, Maluuba’s head of product, said the company hopes to one day use EpiReader to create programs that can read through boring documents like user manuals and customer service documents in order to answer questions for readers.