Story

Cornell CIS Research Gets Facebook Call Out From Founder Mark Zuckerberg

Not only did a Cornell CIS research paper receive the best paper award at CVPR 2017 but also got a call out on Facebook from none other than Facebook founder Mark Zuckerberg. Cornell researchers Gao Huang (Cornell CS), Zhuang Liu (Tsinghua University), Kilian Weinberger (Cornell CS), and Laurens van der Maaten (Facebook AI) were congratulated by Zuckerberg for their award and research work on “Densely Connected Convolutional Networks.”

“One reason I’m so optimistic about AI is that improvements like this research improve systems across so many different fields – from diagnosing diseases to keep us healthy, to improving self-driving cars to keep us safe, and from showing you better content in News Feed to delivering you more relevant search results. Every time we improve our AI methods, all of these systems get better. I’m excited about all the progress here and it’s potential to make the world better,” said Zuckerberg.

The team of researchers Zuckerberg is referring to found that neural networks whose layers are all connected to each other perform better than the current state-of-the-art residual neural networks currently in place. In their paper, the authors introduce the Dense Convolutional Network (DenseNet), which connects each layer to every other layer in a feed-forward fashion.

“This radically changes the information flow in the network,” said lead researcher Weinberger, Cornell computer science professor. “Because every neuron has access to all prior computations, it turns out DenseNets can be far more parameter efficient – an essential feature on modern hardware, where networks are typically limited by the size of the GPU memory.”

What makes this research so relevant to Facebook and its AI team is its accuracy and smaller memory size. “Facebook users upload around 250 million images per day,” said Weinberger. “To utilize the data they have, which is their main asset, Facebook needs some way to do this more efficiently and with higher accuracy. This could translate into cost savings through reduced computational demand and may also unlock new possibilities because these networks make fewer errors.”