The Age of the Neural Nets

Did you ever think that you’d be living in an age where technology would become a vital part of your daily life? From cell phones to laptops, tablets and wearables, technology has revolutionized our everyday life.

Now Nuance, a developer of voice and language products, is applying Neural Nets to advance voice and language technologies that will make it possible for machines to learn and ultimately understand the human language.

What’s a Neural Net? The standard Deep Neural Network (DNN) is unidirectional with information that flows in just one direction. This begins at the input layer and travels through the hidden layers to the output layer. In terms of Machine Learning, these DNNs are of the “feed-forward” type. They work best when all the information needed to learn is available at the same time.

In a recent blog post, entitled “How Many Neural Nets does it take to Catch the big fish in Machine Learning?,” Nils Lenke, researcher, Nuance, wrote, “…we’re just scratching the service with Neural Nets, which are evolving and changing with many different approaches and challenges to solve.”

The teams at Nuance are applying DNNs to advance speech recognition and natural language understanding as part of their mission to better facilitate communication between people and technology.

One of the challenges is context. “Technically speaking, of course, if you waited till the end of an utterance you could make the whole utterance available to a DNN at once and a feed forward network could access all the info it needs to do the recognition job in just one go,” Lenke wrote. The problem is, with dialogue systems, the engines start working right after an utterance begins and tries to keep up with the speaker to quickly offer up a response once the speaker is done talking in order to emulate natural conversation, which is difficult.

This technology is trying to solve these challenges in a number of ways, including leveraging Recurrent Neural Nets that create a sort of ‘memory’ through feedback loops, and creating NeuroCRFs that combine Neural Nets with Conditional Random Fields (CRF) models.

And sometime in the near future, we might be able to have conversations with artificial intelligence devices that don’t leave us in the uncanny valley.