Understanding deep neural networks

In this episode of the Data Show, I speak with Michael Mahoney, a member of RISELab, the International Computer Science Institute, and the Department of Statistics at UC Berkeley. A physicist by training, Mahoney has been at the forefront of many important problems in large-scale data analysis. On the theoretical side, his works spans algorithmic and statistical methods for matrices, graphs, regression, optimization, and related problems. On the applications side, he has contributed to systems used for internet and social media analysis, social network analysis, as well as for a host of applications in the physical and life sciences. Most recently, he has been working on deep neural networks, specifically developing theoretical methods and practical diagnostic tools that should be helpful to practitioners who use deep learning.

Analyzing deep neural networks with WeightWatcher. Image by Michael Mahoney and Charles Martin, used with permission.

We had a great conversation spanning many topics, including:

• The class of problems in big data, machine learning, and data analysis that he has worked on at Yahoo, Stanford, and Berkeley.
• The new UC Berkeley FODA (Foundations of Data Analysis) Institute.
• HAWQ (Hessian AWare Quantization of Neural Networks with Mixed-Precision), a new framework for addressing problems pertaining to model size and inference speed/power in deep learning.
• WeightWatcher: a new open source project for predicting the accuracy of deep neural networks. WeightWatcher stems from a recent series of papers with Charles Martin, of Calculation Consulting.

Related resources:

• “Deep learning at scale: Tools and solutions” – a new tutorial at the Artificial Intelligence Conference in San Jose
• Ameet Talwalker on “How to train and deploy deep learning at scale”
• Greg Diamos on “How big compute is powering the deep learning rocket ship”