The paper presents a novel approach to bridging the gap between generative AI models and cognitive architectures, specifically the Common Model of Cognition (CMC). The key insights are:
The authors propose a "Middle Memory" (MM) interface between the CMC modules and the underlying generative networks. The MM acts as an intermediary, receiving and managing the outputs from the various networks and making them accessible to the CMC's central production system.
The authors model all CMC modules, except procedural and working memory, as "shadow production systems". These shadow productions fire by matching to information in working memory (WM) and MM, allowing them to refine and contextualize the network predictions before passing them to the central production system.
The central production system remains the core decision-making component, but it now receives a richer, contextualized input from the shadow productions, which can draw on both the network predictions and the structured knowledge represented in MM.
The authors propose integrating the learning mechanisms across the shadow productions and the central production system, such that rewards and utility updates are propagated to all relevant components.
This hybrid architecture aims to combine the strengths of generative networks (for tasks like perception, language, and imagination) with the symbolic reasoning and decision-making capabilities of the CMC, resulting in a more robust and human-like cognitive system.
The proposed framework offers a principled way to bridge the gap between connectionist and symbolic approaches, potentially enabling the modeling of a wider range of human cognitive abilities within a unified architecture.
翻譯成其他語言
從原文內容
arxiv.org
深入探究