如何让通用人工智能系统能够“识数”

How Can an Artificial General Intelligence System Acquire“Number Sense”?

“识数”的字面意思就是说,数字表达式的使用者需要知道这些表达式日常实践中的具体意义,如知道“十公里”算不算“路远”,“十万元”算不算“昂贵”。主流人工智能系统虽然都是数码化的,但却未必真正“识数”,因为它们都无法以一种灵活的方式来理解数字表达式在特定的日常语境中究竟意味着什么,并在这种理解的基础上进行复杂的语用推理。为了使得人工智能系统能够“识数”,工作思路应是在非公理化推时系统的平台上来构建数字表征。训练非公理化推理系统使之“识数”的指导原则,就是先教会它表征那些最不抽象的数学概念,再一步步进展到更抽象的概念。而这一教学步骤的安排,与我们教导儿童识数的路线图是有点相似的。在这个面向机器的数字教学法中,我们将以对“逻辑专名”的教学作为对数字表征教学的重要前导。

Number sense literally means some form of awareness of what numbers mean to the users of these numbers in their practical lives.Mainstream AI systems(symbolic AI and connectionism included),being digitalized notwithstanding,do lack number sense in the sense that they all lack the capacities of building pragmatic inferential pathways connecting numerical expressions with non-numerical ones in a context-sensitive manner.The tentative realization of a machine-based number sense will appeal to the Non-Axiomatic Reasoning System(or“NARS”,the adjective form of which is“Narsian”),and it is different from mainstream.AI systems in the sense that it has capacities for learning from its experience,and for delivering its solutions to the given problems even when it is working with insufficient knowledge and limited time budget.The guiding principle for training NARS to acquire number sense is to teach it the least abstract mathematical notions first,then move to more abstract ones step by step,just like how we teach children mathematics.The Narsian representation of the“logical proper names”plays a pivotal role in this machine-oriented pedagogy.