A Unified Structured Framework for AGI: Bridging Cognition and Neuromorphic Computing.

Cognitive modeling and neuromorphic computing are two promising avenues to achieve AGI. However, neither of them has achieved intelligent agents with human-like proficiency so far. One possibility is that the two fields have developed in isolation at different levels, ignoring each other’s complementary features. In this paper, from a graph perspective, we present a framework that bridges the gap through cross-hierarchy structured representation and computation. Combining top-down and bottom-up design methodologies, coherent coordination of cognitive architecture and underlying neural dynamics is realized, where interpretable representation of entities and relations is constructed by hierarchical neuromorphic graph (HNG) via multi-scale projecting and abstraction. An assembly-based graph-oriented spiking message network is dedicatedly developed to conduct reasoning and learning. Evaluation on multi-modal reasoning benchmark indicates that the approach outperforms pure symbolic rule-based and non-neuromorphic baselines. Besides, the framework is flexible and compatible with the mainstream cognitive architectures meanwhile maintaining rich biological fidelity in order for exploiting non-negligible fine-grained mechanisms that are crucial for functionality emerging. Our methodology offers a brand-new guideline for the creation of more intelligent, adaptable, and autonomous systems.