翻译：

Summary of advanced applications of neural network and artificial intelligence

This paper is a comprehensive article, which introduces the concepts and application of artificial intelligence and neural network.This article was originally selected because the neural network was learned in the process of learning machine learning, because the artificial neural network is closer to the human brain in terms of its principles and functional characteristics. It is not a given program that performs operations step by step, but is able to adapt itself to the environment, summarize the rules, perform some operation, recognition or process control. I think there must be many important applications in artificial intelligence, so I chose this article.

The article is a comprehensive article that introduces the concepts and applications of artificial intelligence and neural networks.The artificial neural network is closer to the human brain in terms of its constitution principle and functional characteristics.It is not a given program that performs operations step by step,but can adapt itself to the environment and summarize the rules.Perform an operation, recognition.,or process control.

The first part mainly introduces artificial intelligence.

Artificial intelligence is defined as an artificial object, such as a computer or machine, that shows intelligent behavior capable of dealing with complex problems. And what is intelligence? This involves issues such as consciousness, self, thinking, and so on, and the only intelligence we know is human intelligence.It is a kind of ability to imagine and create memory understanding, pattern recognition, choice, adapt to change and experience learning. The main purpose of artificial intelligence is to make machines behave more like humans, and secondly, to make machines more like humans in the way they solve complex problems but consume less time than humans. Today, artificial intelligence is divided into two parts: strong ai and weak ai.Strong ai means that machines can think on their own, like the scenes in the movies, and even replace humans.The weak ai is the performance of machines that seem to have intelligence, such as playing chess apps, and all the steps it makes to play chess are stored in a computer in advance. The chess app itself doesn't think or plan. How do you know if the machine has intelligent behavior? In 1950, Alan Turing put forward the Turing test, and there is not much explanation for Turing test, because the teacher introduced it in class.

For the origin of artificial intelligence, it is associated with many disciplines, especially philosophy, logic, mathematics, computing, psychology/cognitive science, biological science/neuroscience.

The second part mainly introduces artificial neural network.This section also introduces three small parts, introducing the concept of artificial neural network,the types of neural network learning methods and an important function in artificial neural network-Artificial incentive function.

The first part introduces the concept of artificial neural network. An artificial neural network is a network of processors (neurons) connected, each with a portion of the local storage space (very small).These neurons operate only their own local data and input data (which are entered in one way through links and circuits), and each neuron uses a rule to know the input signal. Output these signals to other neurons, and this calculation of the output data is called an incentive function.

The structure of the neural network generally has three layers, as shown below. The first layer is the input layer, which is used to interact directly with the outside world, and the second layer is the hidden element, which is used to complete the calculation according to the required function. The third layer is the output layer. 

The second part of neural network learning can be divided into three types: supervised learning, unsupervised learning and strengthened learning. In supervisory learning, each instance consists of an input object and an expected output value.Therefore, errors and differences between the expected and actual results of each node on the output layer can be found, which will be used to determine the weight of the network node (according to 

the learning rules). That is, the expected output value on each node is determined by an external teacher.

There are no external teachers in unsupervised learning, so the way of learning is based on clustering, and according to input, the model set is divided into different classes. This kind of learning model can also be called self-organizing mode, typical example is the hebbian learning law and the competition learning law, and unsupervised learning is more important than supervised learning. Because the brain is usually unsupervised.

Strengthening learning is based on unsupervised learning and supervised learning, and in the process of exploration, by exploring the unknown environment while building an environmental model and learning an optimal strategy, Each action corresponds to a reward, and finally gets the greatest reward for data processing.

The third part introduces three kinds of incentive functions. The first threshold function, when the total input is less than the threshold, sets 0, and when the total input value is greater than the threshold, sets 1.

The second is a segmented linear function that can take values between 0,1, depending on the magnification of the linear operation of a region.

The third is the sigmoid function, which can use a range between 0 and 1, but can sometimes take a range from 1 to 1, An example of a sigmoid function is a hyperbolic tangent.

The third part mainly introduces several advanced applications of neural network.

The first application is the computer interface of the human brain based on neural network. The computer interface of human brain is one of the most promising interface technologies between human and machine.BCI is also called the Siwei interface. It is actually a communication channel between the brain and the computer, which allows the signals sent by the brain to interact directly with external activities, such as controlling a cursor, Or the user can enter a phone number by gazing at the keyboard of a display.The interface provides a means of communication between the brain and the interface it wants to control, and the BCI interface makes it possible for a paralyzed person to write a book or control an electric wheelchair. Eg is the best choice to implement BCI, but brainwaves are very weak and there are many kinds of noise.

A signal is obtained from the human brain, then processed, extracted from the features, then classified, and then fed back to the human through the application interface. The number and speed of BCI research has been growing rapidly over the past five years, with no more than six groups studying it in 1995, and at least 20 groups studying BCI now.

The second application is to understand and describe applications in object behavior. Trajectory analysis is one of the core problems in behavior understanding. Trajectory pattern learning can be used to detect anomalies and to predict object trajectory. A model that learns the semantic region by analyzing the trajectory of a moving object in a scene or framework. The first path is encoded to indicate the location of the image and its instantaneous speed.Then the clustering algorithm is applied to classify the tracks according to different spatial and velocity distributions, and in each cluster, the space of the tracks is close and the speed is similar. This class can represent a mode of activity. Based on this orbital cluster, the statistical model of semantic regions in the scene can be obtained by estimating the density and velocity distribution of each activity pattern.The model is based on the combination of vector quantization neural network and neuron types with short-term memory ability. The resulting pedestrian trajectory model will be used to evaluate the new trajectory, predict the future trajectory of the object, randomly generate new trajectory.

The third application is artificial neural network in computer graphics.

Artificial neural network has played a very important role in the image field. The image designer is trying to combine the actual image with the computer generated image to enhance the visualization of the output object. Using thermal sensing technology can produce some of the most authentic images.

 The fourth application is automatic walking robot and underwater robot.

Automatic walking robot is based on the modular concept. The problem of making an automatic walking robot can be disassembled into several functional problems. Breaking down a complex problem into simple, manageable little problems, and the research in this field combines knowledge of biology, mechanics, and information technology, Then develop a dynamic, stable, mobile vehicle using neural network control.The same is true of underwater robots, and underwater machines help salvage operations, prevent pollution, rescue at sea and marine scientific research. So underwater robots have developed a lot over the years.

The fifth application is facial animation.

Face modeling and animation is one of the most difficult tasks in computer graphics, and it is very difficult to turn life into digital form. Use layered b-surface as a base to create facial animation. Neural networks can be used to learn the features of every face expression in an animation sequence.

The sixth is the neural network to strengthen anti-virus technology.

Artificial neural networks and artificial intelligence play an increasingly important role in virus detection, which strengthens the internal functions of anti-virus technology, allowing it to detect and repair all kinds of viruses. For example, IBM's neural network startup detection technology provides additional security by imitating human neurons to learn the difference between infected and uninfected records.Many examples of viruses and non-viruses show that neural networks perform better than traditional hand-adjusted wizard searches for viruses.

The fourth part mainly introduces the application of artificial intelligence.

The first application is data mining and knowledge extraction. Three basic techniques in artificial intelligence are applied, including knowledge expression, and data mining wants to discover patterns of interest from large amounts of data, which can be used in many forms, such as association rules. Decision rules and decision trees.There is also knowledge acquisition and knowledge reasoning, and the pattern found from the data set needs to be verified in different applications.

The second application is the artificial system. The expert system is a subset of artificial intelligence, and the expert system is an artificial intelligence program, which has expert knowledge in specific fields and knows how to use its knowledge to correctly respond to related problems.

The third application is nature and original process NLP. Natural language processing is a subdomain of artificial intelligence. Its goal is to achieve a human-like language processing mechanism. The following picture is a model of NLP.

The fourth application is cyanology. Robotics is part of the field of artificial intelligence.

The fifth application is to apply artificial intelligence to the game. Modern games usually use 3D animation graphics to give people a real feeling. Artificial intelligence in most computer games is not an academic artificial intelligence, but a very close to artificial intelligence technology, which creates an intellectual illusion.Game artificial intelligence includes techniques that combine programming and design practices: path search, neural networks, emotional models, social scenes, finite state machines, rule systems, Decision tree learning and other techniques.

At the end of the paper, some of the problems that researchers are working on are, for example, whether machines are aware of their existence? What does it mean to humans? Will neural networks be completely similar to the human brain and so on.At the end of the paper, some of the problems that researchers are working on are, for example, whether machines are aware of their existence? What does it mean to humans? Will neural networks be completely similar to the human brain and so on.

The subject of this paper is finished. Through the study and reading of this paper, it is found that the computer world can benefit a lot from the neural network method. In the future, artificial intelligence will develop machines and computers that are more complex than we are today, and they may really have simple common sense and have similar human intelligence in some fields. The future development of artificial intelligence may really change our world.

原文名称为《神经网络和人工智能的先进应用的综述》报告

原文：

《神经网络和人工智能的先进应用的综述》报告

   这篇论文整体上是一个综述性的文章，介绍了一些人工智能以及神经网络的概念以及应用的一些领域。当初选这篇文章，也是因为在学习机器学习的过程中学到了神经网络，因为人工神经网络在构成原理和功能特点等方面更加接近于人脑，它不是给定的程序一步一步的执行运算，而是能够自身适应环境、总结规律、完成某种运算、识别或过程控制，觉得肯定在人工智能方面有许多重要的应用，就选了这方面的文章。

     文章一共分为六个部分，第一二部分主要分别介绍了人工智能以及人工神经网络的概念，以及相关的一些机器学习中的储备知识。第三章讨论了神经网络的应用，第四章讨论了人工智能的应用，第五章是结论，第六章提出未来要解决的问题。

第一部分主要介绍了人工智能。

人工智能定义为一个人工制造的物体，（比如说电脑或者机器）展现出了能够处理复杂问题的智能行为。而智能是什么呢？这涉及到比如说意识、自我、思维等等问题，我们唯一了解的智能就是人本身的智能。文中说是一种能够进行想象创造记忆理解、模式识别、选择、适应改变以及经验学习的一种能力。而人工智能的主要目的，在于让机器的行为变得更像人类，其次，让机器在解决复杂问题的方式上更接近于人类但比人类消耗更少的时间。人工智能研究到今天，分为了两个部分强AI和弱AI。强AI就是说机器能够自主思考，像电影里演的那种场景和人类一样甚至要取代人类。而弱AI就是机器表现的他们好像有智能，比如一些下棋的应用，而它做出的所有下棋的步骤都是被人预先存储在电脑里的，下棋应用本身并没有进行思考或者计划。而怎么知道机器是否有智能行为呢，1950年艾伦图灵提出了图灵测试，这里就不多解释图灵测试，因为老师上课介绍过了。

 对于人工智能的起源，它与许多学科都是相联系的，尤其是哲学、逻辑学、数学、计算学、心理学/认知科学、生物科学/神经科学。

第二部分主要介绍了人工神经网络。这个部分又主要介绍了三个小部分，包括人工神经网络的概念，以及神经网络的学习方法的种类，和在人工神经网络中一种重要的函数人工激励函数。

一个人工神经网络就是许多个处理器（神经元）进行联结而构成的网络，每一个神经元上都有一部分本地的存储空间（很小）。这些神经元只操作自己本地的数据和输入数据（通过链接和线路单向输入进来的），每个神经元用某种规则用来对输入信号进行求知，把这些信号输出给其他神经元，这种对于输出数据的计算规则叫做激励函数。

神经网络的结构一般有三层，如下图所示，第一层是输入层，用来与外界进行直接交互，第二层是隐藏元，用来根据所需功能完成计算。第三层是输出层。

  第二小部分神经网络的学习方法可以分为三种：监督学习，无监督学习以及加强学习。在监督学习中，每个实例都是由一个输入对象和一个期望的输出值组成。因此，输出层上的每个节点的预期与实际结果之间的错误和差异能被发现，它们将用来决定网络节点的权重的改变（根据学习规则）。也就是每个节点上的预期的输出值是由一个外部教师来决定的。

在无监督学习中是没有外部教师的，因此这种学习方式是建立在聚类基础上的，根据输入，模型集合会被划分为成不同的类，这种学习模式也可以叫做自组织模式，典型的例子就是hebbian学习法则和竞争学习法则，无监督学习比监督学习更重要，因为大脑通常是进行无监督学习的。

加强学习是建立在无监督学习与监督学习的基础上的，通过对未知环境一边探索一边建立环境模型以及学得一个最优策略，而在探索过程中，每一个动作对应一个奖赏，最后得到一个奖赏最大的方式进行数据处理。

第三小部分介绍了三种激励函数。第一种阈值函数，当总输入小于阈值的时候，置0，当总输入值大于阈值的时候，置1.

第二种是分段线性函数，可以取0，1之间的值，这取决于某个区域的线性操作的放大系数。

第三种是sigmoid函数，这个函数可以使用0~1之间范围，但是有时候也可以取-1~1范围区间，sigmoid函数的一个例子就是双曲正切函数。

第三部分主要介绍了神经网络几个先进的应用。

第一种应用是基于神经网络的人脑计算机接口。人脑计算机接口是在人类与机器之间非常有前景的接口技术之一。BCI也叫做思维机接口。它其实就是大脑与电脑之间的一个通讯通道，使得大脑发出的信号能够直接与外部活动进行交互，例如控制一个光标，或者使用者可以通过凝视一个显示器的键盘来输入一串电话号码。该接口在大脑与想要控制的接口之间提供了一条通讯途径，BCI接口使得一个瘫痪的人去写一本书或控制一个电动轮椅成为可能。EEG是实现BCI的最佳选择，然而脑电波是非常弱的，而且其中存在许多种噪音。因此选取什么样的特征是有用的，如何提取有用的特征，如何抑制噪音等都是非常重要的。神经网络可以被用来将噪音信号从敏感信号中区分出来，提高意识任务分类的准确性。下面就是一个BCI的简单示意图。

从人脑先获取信号，然后经过处理，进行特征的提取，然后进行分类，最后在通过应用接口，对人进行反馈。在过去的五年间BCI研究的数量与速度都在飞速增长，1995年只有不超过6各小组在研究，而现在至少有20个小组在研究BCI。

第二个应用是理解与描述对象行为中的应用。轨迹分析是行为理解问题中的核心问题之一。轨迹模式学习能够被用来检测异常，也能够被用来预测对象轨迹。通过分析场景或框架中移动物体的轨迹来学习语义区域的模型，首先轨迹会被编码，用来表示图像的位置和它的瞬时速度。然后应用聚类算法将轨迹按照不同的空间和速度分布来进行分类，在每一簇中，轨迹的空间上是接近的，速度上是相似的，这一类能代表一种活动模式。基于这个轨迹簇，通过估计每一种活动模式的密度和速度分布，能够得出场景中语义区域上的统计模型。该模型是基于实现矢量量化的神经网络来和具有短期记忆能力的神经元类型的组合，使用分层自组织神经网络来学习和识别轨迹的分布模式。已经生成的行人轨迹模型会被用来评估新轨迹、预测物体未来的轨迹、随机生成新的轨迹。

第三种应用是计算机图形学中的人工神经网络

人工神经网络已经在图像领域扮演了很重要的角色。图像设计者正试图将实际图像与计算机生成的图像进行结合，进而增强输出对象的可视化。使用热传感技术能够生成一些最真实的图像。

第四种应用是自动行走机器人&水下机器人。

自动行走机器人的基础是模块化概念。制作一个自动步行机器人的问题可以被拆解成几组功能上的问题。把一个复杂问题分解成简单的易于处理的小问题，在这个领域的研究结合了生物学，机械学以及信息技术的相关技术的相关知识，进而开发一个使用神经网络控制的动态的稳定的能够移动的交通工具。水下机器人也是如此，并且水下机器有助于打捞作业、预防污染、海上救援和海洋科学考察等等。因此水下机器人这些年得到了很大发展。

第五种应用是面部动画

人脸建模和动画是当前计算机图形学中最困难的任务之一，将生命转化成数字形式，非常难。使用分层B-曲面来作为底层来创建面部动画。神经网络可以被用来学习动画序列中每个人脸表情的特征。

第六种是神经网络加强防病毒技术

人工神经网络与人工智能在病毒检测中扮演者越来越重要的角色，它强化了防病毒技术的内部功能，让其可以检测和修复所有种类的病毒。比如说，IBM的神经网络启动检测技术通过模仿人类神经元来学习被感染与未被感染的记录之间的区别，进而提供了额外的安全性。许多病毒和非病毒的例子都表现出，神经网络学习法比传统手工调整变化的级法师搜索，在识别病毒方面表现的更好。

第四部分主要介绍了人工智能的应用。

第一个应用就是数据挖掘和知识提取。人工智能中的三种基本技术被应用包括知识表达，数据挖掘想要从大量数据中发现感兴趣的模式，这些模式能够使用多种形式，比如关联规则，决策规则和决策树。还有知识获取和知识推理，从数据集中发现的模式需要在不同的应用中进行验证，因此对于挖掘结果的推演是数据挖掘应用程序必备的一项技术。

第二个应用是人工系统。专家系统是人工智能的一个子集，专家系统是一个人工智能程序，它具有特定领域的专家级知识并且知道如何利用其知识来正确响应相关问题。

第三个应用是自然与原处理NLP。自然语言处理是人工智能的一个子领域。它的目标是实现类似人类的语言处理机制。下图是NLP的模型。

第四个应用是机器人学。机器人技术是人工智能领域的一部分。

第五个应用是将人工智能应用于游戏。现代的游戏通常使用3d动画图形来给人一种真实的感觉。大部分电脑游戏中的人工智能并不能算是学术上的人工智能，而是一种很接近人工智能技术，它创造了一种有智力的错觉。游戏人工智能包括编程和设计实践相结合的技术：路径寻找，神经网络、情感模型、社交场景、有限状态机、规则系统、决策树学习和其他一些技术。

文章最后也提出了一些研究人员正在解决的一些问题：比如机器是否能意识到自己的存在？对人类意味着什么？神经网络会不会与人脑完全相似等等。我觉得唯一确定的是，人工智能正在以非常快的速度发展着，它在给我们带来惊喜与方便的同时，也可能会有一些弊端慢慢显露出来，需要人类共同的努力来发展它和控制它。

本篇论文的主体也就完成了。通过这篇论文的学习与阅读，发现其实计算机世界能够从神经网络方法中受益良多。往后人工智能将会开发出比我们现在更加复杂的机器与电脑，它们可能真的会有简单的常识并且在某些专业领域能具有类似人类的智力。未来人工智能的发展也许真的会使我们的世界焕然一新。