摘要
针对深度神经网络 AI研究的可解释性瓶颈,指出刚性逻辑(数理形式逻辑)和二值神经元等价,二值神经网络可转换成逻辑表达式,有强可解释性。深度神经网络一味增加中间层数来拟合大数据,没有适时通过抽象把最小粒度的数据(原子)变成粒度较大的知识(分子),再把较小粒度的知识变成较大粒度的知识,把原有的强可解释性淹没在中间层次的汪洋大海中。要支持多粒度的知识处理,需把刚性逻辑扩张为柔性命题逻辑(命题级数理辩证逻辑),把二值神经元扩张为柔性神经元,才能保持强可解释性。本文详细介绍了从刚性逻辑到柔性逻辑的扩张过程和成果,最后介绍了它们在 AI研究中的应用,这是重新找回 AI研究强可解释性的最佳途径。
In view of the restrictions on the interpretability of artificial intelligence (AI) research on deep neural net-works,it is indicated that rigid logic (mathematical formal logic) and binary neurons are equivalent.Moreover,a binary neural network can be converted into a logical expression,which is highly interpretable.The deep neural network blindly increases the number of intermediate layers to fit big data without the timely abstraction of data with the smal-lest granularity (atom) into knowledge with larger granularity (molecule),changes knowledge with smaller granularity into knowledge with larger granularity,and submerges the original strong explanatory power in the ocean of intermedi-ate layers.To support knowledge processing of multiple granularities,rigid logic should be expanded into flexible pro-positional logic (proposition-level mathematical dialectic logic) and binary neurons should be expanded into flexible neurons to maintain the strong explanatory power.This paper introduces in detail the achievement of the expansion pro-cess from rigid logic to flexible logic and its application in AI research,which is the best method to recover the inter-pretability of AI.
作者
何华灿
HE Huacan(School of Computer Science,Northwestern Polytechnical University,Xi’an 710072,China)
出处
《智能系统学报》
CSCD
北大核心
2019年第3期393-412,共20页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金面上项目(60273087)
西北工业大学基础研究基金重点项目(W18101)
关键词
人工智能
可解释性
演化
不确定性
泛逻辑学
柔性命题逻辑
柔性神经元
数理辩证逻辑
artificial intelligence
interpretability
evolution
uncertainty
universal logic
flexible propositional logic
flexible neurons
mathematical dialectic logic
