Machine learning idol Jeffrey Dean was one of Google’s earliest employees. Since joining Google in 1999, Jeff has invented MapReduce, Bigtable, Spanner, and most recently TensorFlow. All are widely used cloud computing tools for distributed computing, database, and scalable machine learning. Jeff is currently a Google Senior Fellow working on the Google Brain project. He also is a member of National Academy of Engineering, and a Fellow of the ACM.
At summit.cs during his fully-packed distinguished lecture titled Building Intelligent Systems with Large Scale Deep Learning. Jeff reflected on Google's journal in large-scale deep learning research. It all started from their ICML 2012 work on unsupervised learning for cat images. "Now with more computing power and more data, neural networks work better than other methods," said the machine learning guru. "We have observed surge use of deep learning across lots of Google products in recent quarters," he continued.
More specifically, Jeff noted that after the deployment of deep learning models in Google's speech recognition product system, they have observed a 30% reduction in word error rate. Google's own open-source deep learning software TensorFlow now has 475 non-Google contributors and thousands of stars in Github.
Looking into the future, Jeff remarked that "We see huge potentials of deep learning in health analytics (e.g. detecting diabetes via medical imaging)." Deep learning is creating revolutions in many different fields. One example is that Google Inbox’s Smart Reply was an April Fool’s Day joke in 2009; but with deep learning technology it became a reality after 2015.
At the conclusion of his talk, Jeff revealed some new projects that his team is working on, including using deep reinforcement learning for automated machine learning and model-generating models. During that same morning, Jeff also interacted with 50+ undergraduate and graduate students in an hour-long Q&A session.
Photo Credit: Scott Condon