CoCoSys: The Co-Design of Cognitive AI from Algorithms to Systems

Abstract

—The deep-learning scaling that drove the past decade of AI is colliding with hard limits in energy, data, robustness, and explainability, just as emerging applications demand the capabilities neural networks lack: reasoning, abstraction, and collaboration. The CoCoSys JUMP 2.0 center confronts this by co-designing cognitive systems vertically, from algorithms through architectures and silicon. This article highlights its research along four thrusts: (1) unified neural, symbolic, and probabilistic algorithms; (2) algorithm-hardware co-design steered by systematic workload characterization; (3) technology-driven hardware motifs built on memory-centric, non-volatile-memory computing; and (4) collective and collaborative intelligence for embodied, federated, and human-AI systems. Using neuro-symbolic AI as an end-to-end case study, from algorithm restructuring through reconfigurable dataflows to a heterogeneous silicon prototype, we show that co-designing the full software-to-silicon stack yields orders-of-magnitude efficiency gains while preserving the accuracy and interpretability that make cognitive AI compelling. We close with cross-layer lessons and open challenges.