Thursday, April 25, 2019 - 12:15pm to 1:15pm
Location:3001 Newell-Simon Hall
Speaker:MANISH PANDEY, Fellow and Vice President of Engineering /MANISH%20PANDEY
Architectural Exploration using Reinforcement Learning + Managing Career: A Conversation for Graduate Students
This is a two-part talk. I will start with a presentation of some of our recent work on architectural exploration of domain-specific processors, and how we are using reinforcement learning in the search for efficient AI processor architectures.
The end of Moore’s Law and Dennard scaling, and the exploding computational needs for AI is driving the development of domain-specific processors and systems. Over the last few years, there has been an explosion of AI processor designs for the datacenter and the edge for a broad range of applications ranging from speech recognition, self-driving cars and home automation. These applications impose energy and performance constraints on systems, and few tools are available today to automate the design of deep learning processors optimized for specified performance, accuracy and power metrics. The presentation covers our recent research in CMU SV where, using deep reinforcement learning with actor-critic networks, we achieve a 95% reduction in model size, while maintaining accuracy of within 1% of baseline for convolutional neural networks within a specified power profile. Using multi-objective reinforcement learning with detailed memory models and profiling-based power measurements, we create models optimized for both power and accuracy.
The second half of the talk will be closer to a conversational town hall format. I will start with my observations on managing career for graduate students, across different CS domains and different industry functions. Be prepared to ask your burning career-related questions.
Manish Pandey completed his PhD in Computer Science at CMU in 1997. He is currently a Fellow and Vice President of Engineering at Synopsys, Inc. in Mountain View and is an Adjunct Professor at CMU Silicon Valley. He leads the machine learning and formal verification engineering teams at Synopsys. He has extensive experience in machine learning, distributed systems and infrastructure, and he led the display ad targeting and security group at Yahoo!, and storage analytics systems at Nutanix. He previously led the development of several formal verification technologies at Verplex and Cadence that are in widespread use in the industry. At CMU, his research spans the areas of AI architecture development and NLP.