Speech Information:

to be updated.

Speech Information:

Title: An Alternative Way Towards Human-level Intelligence and Beyond

Abstract: What is a viable way of achieving human-level intelligence and beyond? Rather than a huge, monolithic model with an enormous number of parameters (such as most of the current LLMs), this talk argues that a more modular, more human-inspired approach would fare better and would more likely lead to human-level intelligence and beyond. This position may be argued from the viewpoint of computational cost and consequent energy consumption and environmental impact, but it may also be argued based on the many advantages of reverse-engineering the best intelligent system that we currently have --- the human mind/brain. This position may even be argued from a philosophical or a psychological vantage point. In this regard, the notion of rationality is a case in point: Various types of human rationality can be argued to be needed, and all should be captured in any human-level intelligent system. Therefore, dual-process theory can be argued to be necessary to capture all these types of rationality, which leads naturally to neural-symbolic models. This talk will rely on one such example --- a computational cognitive architecture, which is a comprehensive, domain-generic computational model of human psychological processes, painstakingly designed and extensively tested and validated through explaining or predicting psychological and other empirical data. The talk will address how this modular, human-inspired model incorporates various psychological functionalities, through both symbolic and neural (e.g., LLMs) computation and their interaction. Such models may lead to intelligent systems that are human-like in many ways, making developing, interacting with, and coordinating with them more manageable.