认知机器的结构和激活

PDF下载HTML

Title:

Structure and activation of cognitive machines

作者:

李德毅¹, 张天雷², 韩威³, 海丹³, 鲍泓⁴, 高洪波⁵

1. 清华大学 信息科学技术学院, 北京 100084;
2. 北京主线科技有限公司, 北京 100083;
3. 北京中科原动力科技有限公司, 北京 100085;
4. 北京联合大学 机器人学院, 北京 100101;
5. 中国科学技术大学 信息科学技术学院, 安徽 合肥 230026

Author(s):

LI Deyi¹, ZHANG Tianlei², HAN Wei³, HAI Dan³, BAO Hong⁴, GAO Hongbo⁵

1. School of Information Science and Technology, Tsinghua University, Beijing 100084, China;
2. Trunk Technology Corporation Ltd., Beijing 100083, China;
3. AIforce Technology Corporation Ltd., Beijing 100085, China;
4. School of Robotics, Beijing Union University, Beijing 100101, China;
5. School of Information Science and Technology, University of Science and Technology of China, Hefei 230026, China

关键词:

认知核; 智痕元胞网络; 纠缠; 记忆智能; 具身智能

Keywords:

cognitive nucleus; intelligent trace meta-cell networks; entanglement; memory intelligence; embodied intelligence

分类号:

TP18

DOI:

10.11992/tis.202409024

摘要:

从物理学的角度理解人类认知已经成为当今人工智能面临的核心挑战。本文分析了计算机通用结构中，孤立计算、忽视记忆和孤立思维、忽视具身的局限性。以驾驶认知为例，提出了包括感知、思维、行为在内的认知机器的通用结构组成。它区别于计算机的架构，也区别于杨立昆的“世界模型”和李飞飞的“空间模型”，增加了记忆组块，用人工智痕元胞作为神经元细胞的镜像，用智痕元胞网络构成思维软构体，实现记忆智能的生成、调控和提取。物质硬构体可采用 CPU、DPU、GPU、TPU、FPGA、SSD、搜索引擎等并行处理单元，其系统是分布的、并行的、异构的。一旦加电获得能量，机器就不再是死物质，认知核中的思维软构体和物质硬构体经过一番纠缠，机器被激活。激活后的机器赖负熵为生，进入和物理世界具身交互的认知状态。机器认知像人又不像人，宕机后可再激活，能自主感知、思维、决策和行为，可交互，会学习，自纠错，自成长。该结构既可用于构造数字虚拟机器人，也可用于构造替代人类劳动岗位的、千姿百态具身的机器人，使得人类能够迅速进入人机共生的智能时代。

Abstract:

Understanding human cognition from a physical perspective is a core challenge for artificial intelligence (AI). This study analyzes the limitations of isolated computing, neglecting memory and isolated thinking, and overlooking embodied intelligence within the general structure of computers. In the driving cognition scenario, we propose a general structure of cognitive machines encompassing perception, thinking, and behavior, which differs from the computer architecture, as well as from Yann LeCun’s “world model,” and Li Feifei’s “space model.” It introduces memory blocks and employs artificial intelligent trace meta-cells as mirrors of neural cells, and intelligent trace meta-cells networks are used to form the thinking soft-structured ware to realize the generation, regulation, and extraction of memory intelligence. The hard-structured ware can use parallel processing units such as CPU, DPU, GPU, TPU, FPGA, and search engines. Therefore, the system must be heterogeneous. Once powered on, the machine is no longer a pile of inert matter; it will be activated through the entanglement between the thinking soft-structured ware and the material hard-structured ware in the cognitive nucleus. The activated machine thrives on negative entropy and enters a cognitive state of interaction with the physical world. Machine cognition is both human-and non-human-like; it can be reactivated after downtime, achieving independent perception, cognition, decision-making and behavior, interaction, learning, and self-growth. This structure can be used to construct digital virtual robots and intelligent agents with a myriad of shapes to replace human labor, rapidly propelling humanity into an intelligent era of man–machine symbiosis.