Zhang Yu

jenizhang13@163.com

Deep learning has enabled many practical applications of machine learning and by extension the overall field of artificial intelligence in recent years.

Terrence Sejnowski, a U.S. scientist in computational neuroscience and author of “The Deep Learning Revolution,” said the changes that deep learning brings to fields such as image recognition and speech recognition are just beginning to manifest.

He compared the status quo of deep learning research to the historical milestone of James Watt inventing the steam engine.

“We’re at that very first stage today where James Watt was [at the Industrial Revolution]. And now what’s happening is mathematicians are taking a close look and figuring out a theory that will help us make it more efficient and reliable,” Sejnowski told the Shenzhen Daily at the 2019 Global Mobile Internet Conference held in Guangzhou recently.

He said the basic principles of deep learning require lots of examples and data, which is a departure from the traditional way of solving a problem by writing a program.

“You have to know how to solve a problem yourself to write a program. What if you don’t know how to solve the problem? Well, if you have a system that could figure out on its own by giving it examples, then there are many problems you can solve. That’s why deep learning has been very effective,” said Sejnowski.

He added that architecture is also important for deep learning. The architecture of the most popular deep learning network, the convolutional neural network, is based on how the primate visual system is organized and the many layers in its hierarchy.

“Ultimately, when you get to the top, you can label them. In other words, you can distinguish different types of objects like cups and books at that top layer at the top level,” Sejnowski said.

As the reason why deep learning networks work as well as they do is not fully comprehended, he said mathematicians are needed to use their sophisticated tools and techniques to examine the geometry of networks at different levels.