The book "Algebraic Thinking" compares the difference between the multilayer perceptron and the human brain, tells some shortcomings that the multilayer perceptron cannot achieve, and gives the author's corresponding suggestions. Through reading, I understand that the three basic components of cognition are: the representation of the relationship between variables, the structural representation, and the individual representation that is different from the type representation. These three parts are also the core principles of symbol manipulation. Symbol manipulation contains Based on the above three separable hypotheses: thinking represents the abstract relationship between variables, the mind has a recursive structured representation system, and the mind can distinguish between individual mental representations and types of mental representations. It is precisely because the human brain has the above functions that it can carry out complex cognition and symbol manipulation. In this regard, the multilayer perceptron is flawed.
During the first reading process, the author encountered obstacles because I did not understand some basic concepts. After consulting relevant materials, I gradually became clear: the multilayer perceptron is a feedforward artificial neural network model, which will input multiple data sets. Map to a single output data set. Multi-layer perceptron is a multi-layer fully connected neural network model belonging to connectionism. Why did the author choose multi-layer perceptron as the object of comparison among many neural network models belonging to connectionism? I guess, maybe multi-layer perceptron It can provide a real solution to symbol manipulation better than other models.
Here, the concept of connectionism is also questioned. After consulting the information, I initially figured out the difference between connectionism and behaviorism. Connectionism is a psychological theory, which believes that the connection formed between situational feeling and impulse response is the basis of learning and the basic unit of mental behavior. There are two types of connection, one is the innate connection or reaction tendency, that is, instinct, and the other is the learned connection or reaction tendency, that is, habit. The main point of behaviorism is that psychology should not study consciousness, but behavior. It completely opposes behavior and consciousness. In terms of research methods, behaviorism advocates the use of objective experimental methods instead of introspection. How to apply these two psychological theories to human-computer fusion intelligence, how to integrate human intelligence into the intelligence of machines through thinking about people, and how to build a model of self-perception by studying psychology and studying the individual's cognition of the world , And then establish a model of deep situational awareness, perhaps starting from the study of human-computer fusion psychology .
Connectionism has had a profound impact on symbol manipulation. Someone once said that symbolism, connectionism, and behaviorism are basically handled through some kind of symbol. The core basic concept of human-machine integration of intelligence, artificial intelligence, and deep situational awareness may also be symbols. Therefore, the manipulation of symbols is crucial. Symbol manipulation is the paradigm of information processing systems. The symbol manipulation paradigm regards knowledge as a structure for obtaining information so that concepts can be understood. Both animate (human) or artificial (computer) information processing systems are symbolic manipulation. Symbols are patterns, such as language, marks, signs, etc., which have dual attributes: one is the function of characterizing external things; the other is that they also have physical or formal characteristics that can mark the operation of information processing. The symbol manipulation paradigm regards knowledge as the structure of obtaining information, obtaining recognition or constructing the mode of symbols, so as to understand concepts, and show some general abilities such as being able to reason, solve problems, and use and understand language.
The book "Algebraic Thinking" gives a good introduction to the three core principles of symbol manipulation in chapters.
First, thinking represents the abstract relationship between variables.
Thinking has a special mechanism and form of expression, which can enable the brain to express, extract and generalize the abstract relationship between variables, while the multilayer perceptron cannot realize the abstract relationship among variables, so it cannot fully capture certain empirical facts. Humans can freely generalize one-to-one mapping, even infants can freely map, infants can extract a certain abstract structure from habituation, and can freely generalize. But those multi-layer perceptron models that assign multiple nodes to each variable and train using local learning algorithms cannot. The multi-node multi-variable model trained through backpropagation lacks the ability to freely induce abstract relationships. For this situation, we must seek alternative models. The problem of multi-generalization is the important difference between humans and computers. Humans are not computers. People can directly feed back from input without processing. Humans can handle different kinds of variables, and they can process different kinds of variables together across domains. , How to realize the free generalization of one-to-one mapping like human beings may be the breakthrough point of artificial intelligence.
Second, the mind has a system of recursive structured representation.
大脑具有内部表示结构化知识的方式，而多层感知器是不支持这种结构化表示的。递归（递归关系就是实体自己和自己建立关系，也就是在运行的过程中调用自己。）地表达知识片段的能力是人类认知的核心，像标准多层感知器这样的模型很难捕捉我们表示这些知识的能力，对此，作者提出了一种建议，他想将寄存器按层次结构排列成小树，以表示结构化知识，并且将此类设备可以用作代表我们有关个人知识的基础。这个建议对于多层感知器的改进，或许可以在今后的模型中实现对于知识的结构化表示。
第三，头脑可以区分个人的心理表征和类型的心理表征。
人的大脑或者说思维可以分清类型和个体表征之间的区别，而多层感知器是不支持类型和个体表征的区分。多层感知器在追踪个体时存在困难，因为模型的输入只包含可感知的特性，比如说把两个不同位置的特征相同的物体作为输入，纵使一个是虚拟场景的输入一个是真实场景的输入，但物体可以被感知到的特性是相同的，所以不论他们处于什么位置，在模型中他们的输入就是相同的，因此模型无法像大脑一样区分它们，但是实际上它们并不是同一物体。因而多层感知器无法提供一种将个体与类型分开追踪的方法。
多层感知器不使用符号，符号操纵使用符号，那么是否是大脑在操纵符号呢？作者提出了如何在儿童的大脑中发展符号操纵的机制以及如何在整个进化时期塑造这种机制的问题。目前不清楚符号操纵的能力是不是由自然选择产生的，我们不清楚符号操纵机制是否是DNA决定的与生俱来的能力，但是的确操纵符号的能力是人类和许多动物所拥有的，这对于他们的生存有很大的帮助，而弄清楚符号操纵的产生对于人机融合智能的发展有着极其重要的影响。人的符号极其广泛，如何将人类从基因获得的先天的符号定量定性，如何将人类后天学习的符号进行表征，如何把自然语言和符号语言相互混编，融合在机器之中，使机器正确理解人的思想，获得自主性，或许都是从研究符号开始的。
通过阅读本书，我们大概可以对符号操纵机制有所了解，也明白一些当前多层感知器的缺陷。至于如何修正这些缺陷，作者在书中给出了一些建议。可是笔者对于神经网络模型没有深入系统的了解，所以没有办法实现他的这些建议。人们往往是在遇到挫折时才会发现自己的不足，当阅读本书多层感知器部分，会发现了对于神经网络机器学习我们几乎一无所知，希望可以在之后搞清楚这些概念，系统研究一下机器学习与人类学习之间的基本差异，以便于更好的理解多层感知器，进而可以更好地对于人机融合智能以及深度态势感知进行思考。
也许人类的递归是一个更复杂的系统