2018 - Present : With a background in electronic design automation, I am leveraging my expertise in developing algorithms for large graph transformations to design algorithms for synthesizing neural networks. Neural networks and logic networks share many similarities - they are large directed acyclic graphs with each node performing a possibly unique computation, they are designed to minimize a target metric (delay or area in logic networks, accuracy or parameter count in neural networks). However, defining transformations for neural networks pose unique challenges such as convergence of gradient descent solution, initial value assignment. My current research is geared towards defining these transformations and developing a platform for transformation of neural networks.
2011 - 2017 : I designed new algorithms and techniques for logic and high-level synthesis of digital circuits. My contributions resulted in fundamental advancements in the state-of-the-art design automation tools, thus providing differentiating technology for products and services built on IBM systems. Interacting closely with circuit and logic design leads, I identified drawbacks in the in-house state-of-the-art design automation tool, shaped and developed scalable algorithms for logic and high-level synthesis with the primary goal of meeting specified performance and power consumption constraints when applied to the design of modern system architectures. I also recommended solutions and improvements to problems that arise during the design implementation process. Finally, I collaborated with industry experts, academic professionals, and mentored summer students to a create and pursue a strategic agenda for realizing the long-term goal of a fully automated design implementation process.