Events

Exploring Autonomous Intelligence Beyond the Digital von Neumann Abstraction

Artificial intelligence (AI) and machine learning (ML) technologies have fueled many burgeoning applications from on-device learning and personalized recommendations to self-driving cars and collaborative robots. Despite these unprecedented advancements, the holy grail of enabling fully autonomous machine intelligence remains far from our grasp. One key challenge is the lack of performant and efficient hardware implementation, especially on edge devices with stringent resource constraints.

In this talk, I will present our recent efforts to tackle this challenge from a unique angle that explores system/design abstractions beyond the conventional binary digital logic and von Neumann architecture. Specifically, we leverage information processing ability innate in the analog/mixed-signal (AMS) domain and exploit co-located compute and memory organization. Our AMS-domain method not only leads to novel resistive RAM (RRAM) based analog-to-digital converters that exceed the state-of-the-art performance, but it also transforms the design of peripheral circuits in Processing-in-Memory (PIM) systems delivering much-improved computing performance and efficiency. Taking one step further than simply addressing the compute bottleneck, we demonstrate how PIM architecture can accelerate sparse memory-heavy embedding operations and propose a near-memory solution for large-scale personalized recommendation systems.

Besides pushing the envelope of computing both on the edge and in the cloud, I will also briefly highlight the potential of this beyond-digital approach in diverse fields such as learning-assisted design automation, analog/physical domain security, and real-time adversarial cyberphysical systems. I will conclude the talk with a vision for future autonomous intelligence powered by distinctive yet complementary computing paradigms.