9人工智能导论 (3).pdf

《9人工智能导论 (3).pdf》由会员分享，可在线阅读，更多相关《9人工智能导论 (3).pdf（17页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Three Perspectives on Machine Learning Artificial Intelligence 34 9. Perspectives about Machine Learning 9.1. What is Machine Learning 9.2. History of Machine Learning 9.3. Why Different Perspectives 9.4. Three Perspectives on Machine Learning 9.5. Applications and Terminologies Contents: Artificial

2、 Intelligence 35 9.4. Three Perspectives on Machine Learning 9.4.1. Learning Tasks 9.4.2. Learning Paradigms 9.4.3. Learning Models Contents: Artificial Intelligence : Learning : Perspectives 36 The learning tasks are used to denote the general problems that can be solved by learning with desired ou

3、tput. 学习任务用于表示可以用机器学习解决的基本问题。 What are Learning Tasks 什么是学习任务 9.4.1. Learning Tasks Various types of problems arising in applications: 应用中会产生各种类型的问题： computer vision, 计算机视觉， pattern recognition, 模式识别， natural language processing, 自然语言处理， etc. 等等。 Why Study Learning Tasks 为什么要研究学习任务 Artificial Intell

4、igence : Learning : Perspectives 37 Typical Tasks in Machine Learning 机器学习中的典型任务 9.4.1. Learning Tasks Tasks 任务 Brief Statements 简短描述 Typical algorithm 典型算法 Classification 分类 Inputs are divided into two or more known classes. 将输入划分成两个或多个类别。 SVM 支撑向量机 Regression 回归 Outputs are continuous values rathe

5、r than discrete ones. 输出是连续值而不是离散的。 Bayesian linear regression 贝叶斯线性回归 Clustering 聚类 Inputs are divided into groups which are not known beforehand. 输入被划分为若干个事先未知的组。 k-means k-均值 Ranking 排名 Data transformation in which values are replaced by their rank. 用它们的排名来代替值的数据转换。 PageRank 网页排名 Density estimati

6、on 密度估计 Find the distribution of inputs in some space. 寻找某个空间中输入的分布。 Boosting Density Estimation 增强式密度估计 Dimensionality reduction 降维 Simplify inputs by mapping them into a lower dimensional space. 通过将输入映射到低维空间来将其简化。 Isomap 等距特征映射 Optimization 优化 Find the best solution from all feasible solutions 从所有

7、可能的解中寻找最优解。 Q-learning Q-学习 Artificial Intelligence : Learning : Perspectives 38 Case study: Credit scoring 信用评分 9.4.1. Learning Tasks IF income 1AND savings 2 THEN low-risk ELSE high-risk Two classes: Low-risk and high-risk customers. 二分类：低风险和高风险客户。 A customer information makes up the input to one

8、of the two classes. 客户信息使该输入构成二分类中的一个。 After training with past data, a classification rule learned may be: 用过去的数据训练之后，可以学习得到如下分类规则: Artificial Intelligence 39 9.4. Three Perspectives on Machine Learning 9.4.1. Learning Tasks 9.4.2. Learning Paradigms 9.4.3. Learning Models Contents: Artificial Inte

9、lligence : Learning : Perspectives 40 The learning paradigms are used to denote the typical scenarios that are happened in machine learning. 学习范式用于表示机器学习中发生的典型场景。 What are Learning Paradigms 什么是学习范式 9.4.2. Learning Paradigms by the scenarios or styles in machine learning about 根据机器学习的典型场景或样式： how it

10、 learns from data, 它怎样从数据中学习， how it interactives with environment. 它如何同环境互动。 How to Distinguish Learning Paradigms 怎样区分学习范式 Artificial Intelligence : Learning : Perspectives 41 Typical Paradigms in Machine Learning 机器学习中的典型范式 9.4.2. Learning Paradigms Paradigms 范式 Brief Statements 简短描述 Typical Algo

11、rithm 典型算法 Supervised 有监督 The algorithm is trained by a set of labeled data, and makes predictions for all unseen points. 算法采用一组标注好的数据进行训练，再对所有的未知点做出 预测。 Support vector machines 支撑向量机 Unsupervised 无监督 The algorithm exclusively receives unlabeled data, and makes predictions for all unseen points. 算法仅

12、接收未标注的数据，再对所有的未知点做出预测。 k-means k-均值 Reinforcement 强化 The algorithm interacts with environment, and receives an reward for each action. 算法与外部环境交互，每个动作得到一个回报。 Q-learning Q-学习 Artificial Intelligence : Learning : Perspectives 42 Environment Agent Action State Reward Reinforcement Learning强化学习 learning

13、from state and reward, take better action to the environment. 从状态和回报中学习，对环境采取更好的动作。 Case Study: Deep Reinforcement Learning 深度强化学习 9.4.2. Learning Paradigms Deep Q-network (DQN) 深度Q-网络 (DQN) combines CNN with Q-learning (a form of reinforcement learning), 将CNN与Q-学习（一种强化学习的形式）相结合, input is raw pixels

14、 and output is a value function estimating rewards. 输入是原始像素，输出是估计回报的价值函数。 Artificial Intelligence : Learning : Perspectives 43 Feb. 2015, Google DeepMind published Deep Q-Network, the human-level control through deep reinforcement learning. 2015年2月，谷歌DeepMind发表了深度Q-网络，通过深度强化学习达到人类水平的操控。 Case Study:

15、Deep Reinforcement Learning 深度强化学习 9.4.2. Learning Paradigms Tested on classic Atari video games (late-1970s and early-1980s) Source: NATURE, Feb. 2015 At human-level or above (29/49 59.18%) Below human-level (20/49 40.82%) Artificial Intelligence 44 9.4. Three Perspectives on Machine Learning 9.4.1

16、. Learning Tasks 9.4.2. Learning Paradigms 9.4.3. Learning Models Contents: Artificial Intelligence : Learning : Perspectives 45 The learning models are used to denote the approaches that can handle to fulfil a learning task. 学习模型用于表示可以处理完成一个学习任务的方法。 What are Learning Models 什么是学习模型 9.4.3. Learning

17、Models The result of machine learning is heavily dependent on the choice of an approach for solving the learning task. 机器学习的效果在很大程度上取决于解决该学习任务时所选用的方法。 Why Study Learning Models 为什么要研究学习模型 Artificial Intelligence : Learning : Perspectives 46 Models 模型 Brief Statements 简短描述 Sub-models 子模型 Typical Algo

18、rithm 典型算法 Geometric 几何 Use geometric models such as line, plane, distance or manifold to construct learning algorithms. 采用线、面、距离或流行等几何图形模型来构建学习算法。 Line 线Linear Regression线性回归 Plane面SVM 支撑向量机 Distance 距离k-NNk-近邻 Manifold 流行Isomap等距映射 Logical 逻辑 Use logical models to construct learning algorithms. 采用

19、逻辑模型来构建学习算法。 Logic 逻辑Inductive Logic Program.归纳逻辑编程 Rule 规则 Association Rule 相关规则 Networked 网络 Use networked models to construct learning algorithms. 采用网络模式构建机器学习算法。 Shallow 浅层 Perceptron 感知机 Deep 深层CNN卷积神经网络 Probabilistic 概率 Use probabilistic models to denote the conditional dependence between rand

20、om variables. 采用概率模式来表示随机变量之间的条件相关性。 Bayes 贝叶斯 Bayesian Network 贝叶斯网络 Generative 生成Probabilistic Program.概率规划 Statistic 统计 Linear Regression 线性回归 Typical Models for Machine Learning 机器学习的代表性模型 9.4.3. Learning Models Artificial Intelligence : Learning : Perspectives 47 CNN can be designed to recogniz

21、e visual patterns directly from pixel images with minimal preprocessing. CNN可以设计为用最少的预处理来直接从像素图像中识别视觉模式。 Case Study: Convolutional neural network (CNN) 卷积神经网络 9.4.3. Learning Models Source: Input Layer-3 Layer-1 Layer-5 CNN can recognize patterns with extreme variability, and with robustness to dist

22、ortions and simple geometric transformations. CNN可以识别一些特殊情况的模式，如手写字符识别，对扭曲的以及几何形变的手写字符具有 很强的鲁棒性。 LeNet-5, a pioneering 7-level convolutional neural network LeNet-5, 一种开创性的7层卷积神经网络 Artificial Intelligence : Learning : Perspectives 48 Yann LeCun (born near Paris, France, in 1960) 雅恩乐库（1960年生于法国附近） One

23、 of the fathers of Deep Learning, a founding father of CNN. 深度学习的奠基人之一，CNN的创始人。 Professor at NYU, Director of AI Research at Facebook. 美国纽约大学教授、Facebook人工智能实验室主任。 2014 IEEE Neural Network Pioneer Award. 2014年获得IEEE神经网络先锋奖。 Thought as a modern day Descartes. 被认为是当代的笛卡尔。 Yann LeCun and Rene Descartes 雅恩乐库与雷内笛卡尔 9.4.3. Learning Models Yann LeCun (1960-Current)Rene Descartes (1596-1650)