基於表徵、功能、神經達爾文主義的學習模型探討

Abstract

即使是目前最先進的人工智慧技術在歸納能力方面仍然遠不及人類，人類可以利用基本的背景知識從觀察到的事物中學習並歸納出通則。許多關於人類心智的研究都想要解釋人類的學習能力。在許多關於人類心智的假說中，此論文基於表徵主義、功能主義、神經達爾文主義的想法，設計了一個學習模型。本論文探討此學習模型在很少的基本背景知識下學習通則的能力。本論文提出的學習模型分為三個部分，而表徵主義、功能主義、神經達爾文主義分別為這三個部分奠定了基礎。表徵主義為此學習模型提供了表達知識的方式：此模型中所有的知識皆以符號為最小單位； 功能主義為此學習模型提供了結合知識的方式：此模型的知識可以用函數的方式來結合； 神經達爾文主義為此學習模型提供了歸納知識的方式：此模型利用優勝劣敗的演化方式來學習通則。本論文分析了這個學習模型：此學習模型的表達能力和圖靈機(universal Turing machine)相同；此模型的歸納能力是可靠(sound)但不完備(incomplete)的。本論文利用三個實驗來測試這個學習模型的能力，實驗結果說明此學習模型可以在這些問題中歸納出通則。此論文證實了基於這三個關於人類心智假說的學習模型具備了歸納學習的能力； 即使基於很少的背景知識，此學習模型仍然可以從基本的例子中歸納出通則； 這個學型模型的能力可以利用給予額外的知識來提高歸納學習的效率。

Even the best artificial intelligence technology can not compare with humans on the inductive ability. Humans can learn general rules from observed instances based only on bare-bone prior knowledge. Lots of research on the human mind tries to explain the learning ability of humans. Of all these hypotheses about the human mind, this thesis designs a learning model based on the ideas of representationalism, functionalism, and neural Darwinism. This thesis researches on the learning ability of the proposed model when the model is given only bare-bone prior knowledge. The proposed model is composed of three parts, which are respectively founded on representationalism, functionalism, and neural Darwinism. The knowledge representation of the proposed model is designed based on representationalism: symbols are the basic components of the knowledge representation of the model. The combination of knowledge of the proposed model is designed based on functionalism: the knowledge in the model can be combined in the form of functions. The inductive ability of the proposed model is designed based on neural Darwinism: the model exploits evolutionary processes to learn general rules. This thesis analyzes the proposed learning model. The expressive power of the model is the same as the universal Turing machine. The inductive process of the model is sound but incomplete. Three experiments are used to test the learning ability of the proposed model. Results shows that the learning model is able to learn general rules in these experiments. The results shows that the learning model is able to induct general rules from input instances of these experiments. This thesis proves that a learning model presuming representationalism, functionalism, and neural Darwinism is able to do inductive learning. Even based on bare-bone prior knowledge, the learning model still can learn general rules from instances. The learning model can exploit additional analogical knowledge to increase its inductive learning ability.