延伸阅读 / Extended Reading：

从逻辑信息模型，到逻辑信息网络，直至实现通用人工智能

From Logical Information Model, to Logical Information Network, to the realization of Artificial General Intelligence (AGI)

Abstract:

In the 2010s, with the gradual maturity of connectionism tools such as deep learning algorithms and neural networks, another wave of artificial intelligence (AI) has been set off. However, it has been more than 6 years since AlphaGo made great achievements in 2016, and AI research has once again encountered a bottleneck on the way to further development. Due to its relying too much on the theoretical basis based on probability theory, connectionism has been unsatisfactory in the aspects of causal reasoning and interpretability, which has been criticized by many people as not being real intelligence.

After analyzing the logic basis of the mainstream schools such as connectionism and symbolism, this article combines the core elements of them and proposes a logical information model with strong logicality and practicability. The model and the large-scale information network based on it are expected to fundamentally improve the problems of causal reasoning and interpretability, and become a new theoretical breakthrough in constructing a new knowledge graph structure, and even moving towards artificial general intelligence (AGI).

However, the theoretical research of this model is still at an early stage. The current model design is only limited to pure text information, and the compatibility of other information sources, such as graphics and sound, remains to be explored. Meanwhile, it's also necessary to test the complex characteristics for the model in the huge data environment of logical information network. Meanwhile, it also awaits to test the complex characteristics for the model in the huge data environment of logical information network. In addition, for trustworthy AI that is required to realize strong logicality, or AGI suitable for different application scenarios, it is also necessary to further develop and verify its core "dynamic information algorithm" in the logical information network environment of massive data.

Although it is still far from the goal of AGI, and there are many problems to be solved, it is still worth trying to summarize and explore new routes when the existing AI routes encounter bottlenecks.

Keywords: Artificial General Intelligence (AGI); Logic; Logical Information Model; Logical Information Network

It involves:

It is closer to traditional logic, and only absorbs some basic concepts in the domain of logic as its core. It is fully compatible with mathematical logic.

It is logicism but not symbolism, or connectionism. It is closer to symbolism, and can be used as the data infrastructure of the new generation of knowledge base or knowledge repository and combined with expert systems. However, it can also be combined with connectionism, and become the next research object of graph neural network because of its intrinsic properties of graph structure.

It is a new kind of information model which combines the most basic elements of logic. Meanwhile, the infrastructure of this model is also based on graph structure. To some extent, it can be regarded as the upgraded form of knowledge graph model (or semantic network model).

1. Introduction

2. Logical Model

• 2.1 Argument-Proposition Model
  • 2.1.1 Argumentation & Proposition
  • 2.1.2 Naive Argument-Proposition Model
  • 2.1.3 Complex Argument-Proposition Model
• 2.2 Definition & Narration
  • 2.2.1 Definition
  • 2.2.2 Narration
• 2.3 Truth-Value of Proposition & Validity of Argument
  • 2.3.1 Truth-Value of Proposition
  • 2.3.2 (Special) Validity of Argument
• 2.4 Condition
  • 2.4.1 Condition of Proposition
  • 2.4.2 Condition of Argument
• 2.5 Reconsider Concept
  • 2.5.1 Concept & Definition
  • 2.5.2 Concept & Proposition

3. Information Model

• 3.1 Doubt & Explanation
• 3.2 Proposer & Observer
  • 3.2.1 Proposer & Observer
  • 3.2.2 Declaration of Proposer & Response of Observer
• 3.3 Information Unit
  • 3.3.1 Information Unit
  • 3.3.2 Atomicity
  • 3.3.3 Assembling & Disassembling
• 3.4 Information Relation
  • 3.4.1 Mapping Relation
  • 3.4.2 Equivalence Relation
  • 3.4.3 Causality Relation
  • 3.4.4 Defining Relation
  • 3.4.5 Other Relations
• 3.5 Information Section
  • 3.5.1 Fusing
  • 3.5.2 Information Section
  • 3.5.3 Separating

4. Logical Information Model

• 4.1 Completeness
  • 4.1.1 Completeness Extension of Statement
• 4.2 Correspondence
  • 4.2.1 Correspondence between Conceptual Connotation and Words
  • 4.2.2 Correspondence between the Keys and Values of Attribute
• 4.3 Conversion
  • 4.3.1 Serialization
    • 4.3.1.1 Serialization of Argument
    • 4.3.1.2 Serialization of Doubt
    • 4.3.1.3 Other Serializations
    • 4.3.1.4 Deserialization
  • 4.3.2 Symbolization
    • 4.3.2.1 Symbolization of Truth-Value of Proposition
    • 4.3.2.2 Symbolization of Validity of Argument
  • 4.3.3 Conversion between D&E and Inference

5. Logical Information Network: Characteristics

• 5.1 Fragmentization
• 5.2 Hierarchicalization
• 5.3 Networking
• 5.4 Quantifiability
  • 5.4.1 Quantifiability
  • 5.4.2 Redundancy
  • 5.4.3 Knowledge Measurement
• 5.5 Visualization

6. Logical Information Network: Advanced Capabilities

• 6.1 Logic Capabilities
  • 6.1.1 Logicality
  • 6.1.2 The Risk of Reasoning Paradox in Scientific Axiom System
  • 6.1.3 (General) Validity of Argument
  • 6.1.4 Expanding & Collapsing
• 6.2 Information Capabilities
  • 6.2.1 Static Information
  • 6.2.2 Dynamic Information
  • 6.2.3 Hierarchy Dividing & Classifying of Domain Information
• 6.3 Act

7. Logical Information Network: Significance

• 7.1 Scientific Research & Popularization
  • 7.1.1 Scientific modeling
  • 7.1.2 Interdisciplinary Research
  • 7.1.3 Science Popularization
• 7.2 Cognition
  • 7.2.1 Reading
  • 7.2.2 Memory
  • 7.2.3 Artificial Intelligence

8. Logical Information Network: Unsolved Problems

9. Summary

10. Acknowledgments

Appendix A. The view design scheme of logical information model

Appendix B. References

From Chapter 2 to Chapter 5 is the basic part of the model, from shallow to deep; Chapter 6 "Advanced Capabilities" is the core of the whole article, in which the most important ideas I want to express in this article are basically reflected; Chapter 7 is the prospect of applications and some ideal application scenarios.

摘要：

进入21世纪10年代，随着深度学习算法、神经网络等连接主义工具的日渐成熟，再一次掀起了人工智能的浪潮。然而，自2016年AlphaGo大放异彩至今已有6年有余，人工智能研究在更进一步的道路上再次遭遇了瓶颈。由于过于依赖基于概率论的理论基础，连接主义在因果推理、可解释性等方面的表现一直不尽如人意，为不少人所诟病还不是真正的智能。

本文在剖析了连接主义以及符号主义等主流学派的逻辑学基础之后，糅合了其中的核心要素，提出具有极强逻辑性和可实施性的逻辑信息模型。该模型以及以此为基础构建起来的大规模信息网络有望从根本上改善因果推理及可解释性等方面的问题，并成为构建全新知识图结构、甚至是迈向通用人工智能的全新理论突破。

不过目前该模型的理论研究尚处于较早期阶段，当前的模型设计仅仅限于纯文字信息，针对图形、声音等其它多种信息来源的兼容还有待探索。同时，该模型也有待于检验在逻辑信息网络的巨量数据环境中所表现出来的复杂特性。另外，对于亟待实现强逻辑性的可信人工智能，或是普适于不同应用场景的通用人工智能来说，也都需要进一步在海量数据的逻辑信息网络环境中来发展和验证其最核心的“动态信息算法”。

虽然距离通用人工智能的目标还非常遥远，有诸多问题亟待解决，但是在现有人工智能路线遭遇瓶颈之时，总结探索新的路线依然值得尝试。

关键词： 通用人工智能 逻辑学 逻辑信息模型 逻辑信息网络

它涉及：

更接近传统逻辑，且仅汲取部分逻辑学领域的基本概念为核心； 完全可以兼容数理逻辑；

逻辑主义，但不是符号主义，也不是连接主义； 更接近于符号主义，它可以作为新一代知识库的底层数据结构，与专家系统相结合； 但也可以与连接主义结合，因其本质上图结构的属性而成为图神经网络下一步的研究对象。

一种全新的结合了一众逻辑学最基本要素的数据模型； 同时这种模型的底层还基于图结构，在某种程度上可以看作是知识图谱模型（或者说是语义网络模型）的升级形态。

一、引言

二、逻辑模型

• 2.1 “论证-命题”模型
  • 2.1.1 论证与命题
  • 2.1.2 “论证-命题”简单模型
  • 2.1.3 “论证-命题”复杂模型
• 2.2 定义与叙述
  • 2.2.1 定义
  • 2.2.2 叙述
• 2.3 命题真值与论证有效性
  • 2.3.1 命题的真值
  • 2.3.2 论证的（狭义）有效性
• 2.4 条件
  • 2.4.1 命题的条件
  • 2.4.2 论证的条件
• 2.5 再谈概念
  • 2.5.1 概念与定义
  • 2.5.2 概念与命题

三、信息模型

• 3.1 疑问与解答
• 3.2 发起者与观察者
  • 3.2.1 发起者与观察者
  • 3.2.2 发起者申明与观察者响应
• 3.3 信息元
  • 3.3.1 信息元
  • 3.3.2 原子性
  • 3.3.3 组合与拆解
• 3.4 信息关系
  • 3.4.1 映射关系
  • 3.4.2 等价关系
  • 3.4.3 因果关系
  • 3.4.4 定义关系
  • 3.4.5 其它关系
• 3.5 信息段落
  • 3.5.1 融合
  • 3.5.2 信息段落
  • 3.5.3 分解

四、逻辑信息模型

• 4.1 完备
  • 4.1.1 陈述语句的完备拓展
• 4.2 对应
  • 4.2.1 概念内涵与语词的对应
  • 4.2.2 属性键值的对应
• 4.3 转化
  • 4.3.1 序列化
    • 4.3.1.1 论证的序列化
    • 4.3.1.2 疑问的序列化
    • 4.3.1.3 其它序列化
    • 4.3.1.4 反序列化
  • 4.3.2 符号化
    • 4.3.2.1 命题真值的符号化
    • 4.3.2.2 论证有效性的符号化
  • 4.3.3 问答推论转化

五、逻辑信息网络·特征

• 5.1 碎片化
• 5.2 层级化
• 5.3 网络化
• 5.4 可量化
  • 5.4.1 可量化
  • 5.4.2 冗余
  • 5.4.3 知识度量
• 5.5 可视化

六、逻辑信息网络·复杂功能

• 6.1 逻辑功能
  • 6.1.1 逻辑性
  • 6.1.2 科学公理体系中的推理悖论风险
  • 6.1.3 论证（广义）有效性
  • 6.1.4 展开与折叠
• 6.2 信息功能
  • 6.2.1 静态信息
  • 6.2.2 动态信息
  • 6.2.3 领域信息分层与分级
• 6.3 行为

七、逻辑信息网络·意义

• 7.1 科学研究与普及
  • 7.1.1 科学建模
  • 7.1.2 跨领域研究
  • 7.1.3 科学普及
• 7.2 认知
  • 7.2.1 阅读
  • 7.2.2 记忆
  • 7.2.3 人工智能

八、逻辑信息网络·待解问题

九、总结

十、致谢

附录一、模型视图设计方案

附录二、参考文献

其中，从第二章到第五章是模型的基础部分，由浅及深； 第六章“复杂功能”是整个文章的核心部分，作者在文中想要表达的最重要的一些思想基本上都体现在这一章的内容里； 第七章则是对于应用的展望，一些比较理想化的应用场景等等。