"Technological advancements have led to a proliferation of machine learning systems to assist humans in a wide range of tasks," writes Professor R. Iris Bahar of Brown CS. "However, we are still far from accurate, reliable, and resource-efficient operations for many of these systems. Despite the strengths of convolutional neural networks (CNNs) for object recognition, these discriminative techniques have several shortcomings that leave them vulnerable to exploitation from adversaries. In addition, the computational cost incurred to train these discriminative models can be quite significant."
Earlier this month, Iris gave a keynote address ("Scalable ML Architectures for Real-time Energy-efficient Computing") at the 32nd IEEE International Conference on Application-specific Systems Architectures and Processors (ASAP 2021), which covers the theory and practice of application-specific systems, architectures, and processors. In contrast with discriminative techniques, Iris explains, discriminative-generative approaches, which combine inference by deep learning with sampling and probabilistic inference models, offer a promising avenue for robust perception and action.
"In my talk," she tells us, "I present our work on Generative Robust Inference and Perception (GRIP), a 2-stage approach for pose estimation that uses a CNN for object detection in the first stage followed by Monte-Carlo sampling in the second. It allows the autonomous system to more thoroughly explore a scene and thereby be less likely to be confused by challenging conditions such as occultion, limited lighting, and shadows that naturally occur in real-world environments. Our GRIP process improved pose estimation accuracy by about 25-40% compared to end-to-end neural network approaches. This was a great result, but it came at the cost of slow run times and high energy consumption when running on a GPU. So, our second challenge was to accelerate the Monte-Carlo sampling process by implementing it directly in customized hardware using commercially available reconfigurable logic, or field programmable gate arrays (FPGAs). Our novel implementation of the algorithm directly in hardware allowed us to estimate object poses in real time, using only 2% of the energy that the GPU required. In the future I would like to see more exploration of hybrid approaches that couple neural networks with other ML techniques implemented directly in hardware. This may allow for less complex neural network architectures and training processes while providing efficient and more robust operation of automous systems."
Iris's research interests lie broadly in the areas of computer system design and electronic design automation. In particular, her research focuses on energy-efficient and reliable computing, from the system level to device level. Most recently, she's explored applications for near-data processing and design of robust machine learning techniques for robot scene perception. She is a recipient of the National Science Foundation CAREER award, the Marie R. Pistilli Women in Engineering Achievement Award and the Brown University School of Engineering Award for Excellence in Teaching in Engineering.
For more information, click the link that follows to contact Brown CS Communications Outreach Specialist Jesse C. Polhemus.