机器学习基石 - Types of Learning
2018-03-09 15:55
447 查看
机器学习基石上 (Machine Learning Foundations)—Mathematical Foundations
Hsuan-Tien Lin, 林轩田,副教授 (Associate Professor),资讯工程学系 (Computer Science and Information Engineering)
Multiclass Classification 输出多个类别 (是非题变为选择题)
Regression (回归分析): deeply studied in statistics
Structured Learning 结构化学习 (huge multiclass classification problem)
Supervised Learning: every xnxn comes with corresponding ynyn
每个输入都知道对应的正确输出
Unsupervised: Learning without ynyn
clustering (聚类)
density estimation (密度分析)
outlier detection (离群点分析)
Semi-supervised: Learn with some ynyn
leverage unlabeled data to avoid ‘expensive’ labeling
标记的代价可能很大,只能做一部分
惩罚错误的行为,奖励正确的行为
例子
learn with partial/implicit information (often sequentially) (一个一个的来学习)
Batch Learning: 喂给机器一批一批的数据 (duck feeding)
监督学习、非监督学习
Online Learning: 一个一个的来 (sequentially)
PLA、增强学习
Active Learning: 主动的去学习 (ask questions)
query the ynyn of the chosen xnxn
improve hypothesis with fewer labels (hopefully) by asking questions strategically
机器无法识别时,让机器自己来问,需要标记的可能就少一些
总结
具体的东西,可以计算,预先有人类智慧的加工
the easy ones for ML
like image pixels, speech signal, etc.
often need human or machines to convert to concrete ones
feature engineer (特征工程) 提取出哪些特征给机器比较好
no physical meaning
need feature conversion/extraction/construction
例如只有一些编号的对应关系
Hsuan-Tien Lin, 林轩田,副教授 (Associate Professor),资讯工程学系 (Computer Science and Information Engineering)
Types of Learning
Learning with Different Output Space
Binary Classification (二元分类)Multiclass Classification 输出多个类别 (是非题变为选择题)
Regression (回归分析): deeply studied in statistics
Structured Learning 结构化学习 (huge multiclass classification problem)
Learning with Different Data Label ynyn
关于监督学习和非监督学习
Multiclass Classification 下的几种情况Supervised Learning: every xnxn comes with corresponding ynyn
每个输入都知道对应的正确输出
Unsupervised: Learning without ynyn
clustering (聚类)
density estimation (密度分析)
outlier detection (离群点分析)
Semi-supervised: Learn with some ynyn
leverage unlabeled data to avoid ‘expensive’ labeling
标记的代价可能很大,只能做一部分
Reinforcement Learning (增强学习)
a very different but natural way of learning惩罚错误的行为,奖励正确的行为
例子
learn with partial/implicit information (often sequentially) (一个一个的来学习)
Learning with Different Protocol f⇒(xn,yn)f⇒(xn,yn)
和机器的沟通方式Batch Learning: 喂给机器一批一批的数据 (duck feeding)
监督学习、非监督学习
Online Learning: 一个一个的来 (sequentially)
PLA、增强学习
Active Learning: 主动的去学习 (ask questions)
query the ynyn of the chosen xnxn
improve hypothesis with fewer labels (hopefully) by asking questions strategically
机器无法识别时,让机器自己来问,需要标记的可能就少一些
总结
Learning with Different Input Space XX
Concrete Features
each dimension of X⊆RdX⊆Rd represents sophisticated physical meaning具体的东西,可以计算,预先有人类智慧的加工
the easy ones for ML
Raw Features
更为抽象,包含很多细节 simple physical meaninglike image pixels, speech signal, etc.
often need human or machines to convert to concrete ones
feature engineer (特征工程) 提取出哪些特征给机器比较好
Abstract Features
需要机器自己去学到特征no physical meaning
need feature conversion/extraction/construction
例如只有一些编号的对应关系
思考题
相关文章推荐
- 3. 机器学习基石-When can Machine Learn? - Types of Learning
- 机器学习基石-3-Types of Learning
- 机器学习基石-Types of Learning
- 机器学习基石笔记 Lecture 3 - Types of Learning
- 机器学习基石 - Feasibility of Learning
- 机器学习演算法 第三讲 Types of Learning——学习笔记
- 4. 机器学习基石-When can Machine Learn? - Feasible of Learning
- ML基石_3_TypesOfLearning
- 【Feasibility of Learning】林轩田机器学习基石
- 台湾大学林轩田机器学习基石课程学习笔记3 -- Types of Learning
- 机器学习学习-Types of learning
- 机器学习基石第四讲:feasibility of learning
- 【笔记】机器学习基石(三)type of learning
- 机器学习中的学习方式-Types of learning
- 机器学习基石-04-1-Learning is Impossible
- 机器学习基石-1-The Learning Problem
- 机器学习基石-04-4-Connection to Real Learning
- 机器学习基石第一讲 The Learning Problem
- 机器学习基石 3.4 Learning with Different Input Space
- 机器学习基石 4.4 Connection to Real Learning