While we have figured out how to teach machines to recognize cats and dogs, the next big challenge in AI is whether machines can make critical decisions like humans do, or may be even better. This ability requires that machines learn from past experiences efficiently like humans do. Humans are entrusted with billion dollar decisions, for example, whether to invest or not to invest, and what to invest in. Can we develop AI that would allow us to trust a machine to make these decisions? The benefits are enormous, we could bring the wisdom of experienced people to those who are not as experienced. This can benefit society in education, healthcare and engineering design. Engineering design (e.g., for SoCs) inherently involves trading off competing objectives, for example, increasing performance while reducing power and cost. Human beings learn over time how to make these tradeoffs but still have to go through many design iterations to arrive at a cost-effective solution. As product complexities increase, the number of iterations goes up increasing design cost. The limited number of experienced engineers limits productivity. Design decisions in the chip business, in particular here at Qualcomm, can make a difference of several hundred million dollars because of the very volume of our market – smartphones. At Qualcomm, we are focusing on teaching machines to make these design trade-offs like a human being, but in much shorter time. Thus we can potentially save design cost and produce more designs with the same number of designers. This branch of machine learning is called reinforcement learning and at Qualcomm we are investing in R&D in this area. Developing Reinforcement Learning algorithms for usecases like SoC design is challenging, because the machine needs to go through several experiences to learn like a human. Creating realistic experiences can take a long time and be impractical, especially in our chip design problem. To this end, we are collaborating with Stanford AI research group to develop new reinforcement learning techniques that can learn from past human experiences without the need to generate new experiences. In this talk we will share some results with reinforcement learning in SoC engineering and demonstrate how it can reduce design costs by making design decisions like a human designer but in significantly shorter time.
Shankar Sadasivam (IEEE M’06–SM’18) received the B.Tech. degree in Electrical Engineering from the Indian Institute of Technology (IIT), Madras, in 2005, and the M.S. and Ph.D. degrees in Electrical and Computer Engineering from the University of Illinois at Urbana-Champaign (UIUC), in 2008 and 2011, respectively. His PhD dissertation focused on developing Machine Learning approaches to traditional signal processing problems in information hiding. Since 2011, he has been with Qualcomm where he has been driving various Machine Learning research and product initiatives, e.g., the Always-On Mobile Contextual Awareness project, where his team designed novel Machine Learning algorithms and architectures for always-On contextual awareness usecases. Currently, he drives Machine Learning R&D to maintain Qualcomm’s leadership in cost, features and power for mobile chipset platforms across multiple product lines, in the face of exponentially increasing product complexity with deep-submicron Moore's law scaling.