A Vertical-Horizontal Integrated Neuro-Symbolic Framework Towards Artificial General Intelligence.

Neuro-symbolic technologies with vertical and horizontal approaches are important for the development of Artificial General Intelligence (AGI). But most of the neuro-symbolic works aim at narrow AI problems and do not have a guideline for AGI. The integration of the two approaches could in principle provide a more holistic framework for AGI research. To our best knowledge, such integration has not been explicitly reported yet. In this paper, we identify that vertical and horizontal neuro-symbolic approaches have independent benefits for investigating AGI problems. We then introduce a framework integrating the two approaches, make the first step to implement it, and discuss future updates. The version-one framework contains a central Spiking Reasoning Network (SRN) and several peripheral perceptual modules. The SRN is a programmable spiking neural network that can do logical reasoning under instructions. The version-one framework is implemented on two visual query answering tasks to investigate the programmability of the SRN and to examine the feasibility of the framework. We also discuss the learnability, the biological plausibility, and the future development of the SRN.