HackerNews中文版

概念化在通用人工智能（AGI）的发展中起着至关重要的作用，因为它将使超级智能能够像人类一样理解和处理抽象信息，但…… 从神经元的角度来看，什么是概念？概念可以定义为一种潜在的、抽象的、多模态的表征，它整合了来自不同来源的感官信息：视觉、听觉、嗅觉、触觉，这些信息被编码在一个超高维空间中，其结构源于多个神经元的激活模式。这种表征是组合式的，因为它由更简单的表征的层次组合构成，并且是关系性的，因为它在拓扑上与相似或功能相关的概念相连，这取决于网络的训练。类似于语言嵌入，概念在潜在空间中的“距离”反映了它们的相似程度：它们越接近，在意义、结构或功能方面就越共享。许多概念与语言标签相关联，但这并非绝对必要；有很多事情我们无法命名。另一方面，语言标签——词语——本身就是概念，也被编码在该神经网络的潜在向量空间中。情况只能如此，因为语言是在同一个网络中处理的。换句话说，我们的概念并非纯粹是语言性的。它们基于直接的经验：我们在现实世界中看到、听到、触摸、闻到、想象或体验到的东西。语言作为一种工具，用于指代和分享这些概念，但它本身并不能定义它们。概念性思维的基础是多模态的，不仅仅是语言性的，而词语只是一种模态，一种元模态。有趣的是，我们对语言进行了深入的研究，甚至创建了一门专门研究其意义的学科：语义学。然而，我们对概念化的过程关注甚少。事实上，我们甚至没有一个词来命名这门学科：我们没有“概念学”或类似的词。所有逃避语言领域的东西——那些不容易简化为词语或方程式的东西——似乎像水一样从我们手中溜走，难以捕捉，难以理解和分析。直到人工智能神经网络的出现，我们才开始理解概念化。

查看原文

Conceptualization plays a crucial role in the development of Artificial General Intelligence (AGI), as it will enable that super-intelligence to understand and handle abstract information in a manner similar to how humans do, but...What is a concept from a neuronal point of view?A concept can be defined as a latent, abstract, and multi-modal representation that integrates sensory information from different sources: sight, hearing, smell, touch, encoded in a hyper-dimensional space, whose structure emerges from the activation pattern of multiple neurons.This representation is compositional, in the sense that it is formed by hierarchical combinations of simpler representations, and relational, as it is topologically connected to similar or functionally related concepts, depending on the network's training.Similar to linguistic embedding, the "distance" between concepts in latent space reflects their degree of similarity: the closer they are, the more they share in terms of meaning, structure, or function.Many concepts are associated with a linguistic label, but this is not strictly necessary; there are many things we cannot name. On the other hand, linguistic labels — words — are themselves concepts that are also encoded in that neural network's latent vector space. It could not be otherwise, since language is processed in the same network.In other words, our concepts are not purely linguistic. They are based on direct experience: what we see, hear, touch, smell, imagine, or experience in the real world. Language serves as a tool to refer to and share these concepts, but it does not define them by itself. The foundation of conceptual thinking is multi-modal, not merely verbal, and words are just a modality, a meta-modality.Interestingly, we have studied language in great depth, to the point of creating an entire discipline dedicated to its meaning: semantics. However, we have paid very little attention to the process of conceptualization. In fact, we do not even have a word to name that discipline: we do not have a "conceptmatics" or equivalent. Everything that escapes the linguistic realm — that cannot be easily reduced to words or equations — seems to slip through our fingers like water, difficult to catch, difficult to understand and analyze. It is only with the advent of artificial neural networks that we have begun to understand conceptualization.

只需要概念化