03 Types of Learning

Output Space/Data Label/Protocol/Input Space四个维度介绍常见机器学习类型,见详细课件

Output Space

的维度考虑,不同的输出空间,对应不同的机器学习算法。

Binary Classification

二分类问题,输出空间为

。常见例子比如:

  • credit approve/disapprove
  • email spam/non-spam
  • patient sick/not sick
  • ad profitable/not profitable

是极其重要的一类问题:

Core and important problem with many tools as building block of other tools.

Multiclass Classification

多分类问题,输出空间为

,二分类是

时候的特例。常见例子比如:

  • coin recognition
  • written digits ⇒ 0, 1, · · · , 9
  • pictures ⇒ apple, orange, strawberry
  • emails ⇒ spam, primary, social, promotion, update
Regression

回归问题,输出空间

或者

,对应bounded regression。常见的例子比如:

  • patient features ⇒ how many days before recovery
  • company data ⇒ stock price
  • climate data ⇒ temperature

统计学中被广泛研究:

Also core and important with many ‘statistical’ tools as building block of other tools.

Structured Learning

结构化学习,常见例子比如:

  • sentence ⇒ structure (class of each word)(序列标注)
  • protein data ⇒ protein folding
  • speech data ⇒ speech parse tree

Huge multiclass classification problem (structure = hyperclass) without ‘explicit’ class definition.

Data Label

从data label

的有无、多少、形式划分:

  • supervised: all
  • unsupervised: no
  • semi-supervised: some
  • reinforcement: implicit

by goodness

Supervised Learning

Supervised learning: every

comes with corresponding

.

比如二分类、多分类问题,都是典型的监督学习。

Unsupervised Learning

Unsupervised learning: diverse, with possibly very different performance goals.

无监督学习形式也很丰富,常见的比如:

  • clustering
    • unsupervised multiclass classification
    • i.e. articles ⇒ topics
  • density estimation
    • unsupervised bounded regression
    • traffic reports with location ⇒ dangerous areas
  • outlier detection
    • extreme ‘unsupervised binary classification’
    • i.e. Internet logs ⇒ intrusion alert
Semi-supervised Learning

Semi-supervised learning: leverage unlabeled data to avoid ‘expensive’ labeling.

常见的比如:

  • face images with a few labeled ⇒ face identifier (Facebook)
  • medicine data with a few labeled ⇒ medicine effect predictor

详细解释见Semi-supervised learning

Reinforcement Learning

Reinforcement: learn with ‘partial/implicit information’ (often sequentially).

样本形式

常见的比如:

  • (customer, ad choice, ad click earning) ⇒ ad system
  • (cards, strategy, winning amount) ⇒ black jack agent

Different Protocol

不同Protocol对应不同Learning Philosophy:

  • batch: duck feeding
  • online: passive sequential
  • active: question asking (sequentially)(query the

of the chosen

)

对应的训练数据也不相同:

  • batch: all known data
  • online: sequential (passive) data
  • active: strategically-observed data
Batch Learning

一次性从所有已知数据中学习。

Batch supervised multiclass classification: learn from all known data.

  • batch of (email, spam?) ⇒ spam filter
  • batch of (patient, cancer) ⇒ cancer classifier
  • batch of patient data ⇒ group of patients
Online Learning

序列地接受数据,然后更新模型。

Online: hypothesis ‘improves’ through receiving data instances sequentially

比如online spam filter, which sequentially:

  1. observe an email
  1. predict spam status with current
  1. receive ‘desired label’

from user, and then update

with

PLA can be easily adapted to online protocol.

Active Learning

当模型没有把握的时候,把问题交给用户,从而获取高质量样本。

Active: improve hypothesis with fewer labels (hopefully) by asking questions strategically

Different Input Space

根据输入空间的含义划分。

Concrete Features

Concrete features: each dimension of

represents ‘sophisticated physical meaning’.

常见的比如:

  • (size, mass) for coin classification
  • customer info for credit approval
  • patient info for cancer diagnosis
  • often including human intelligence on the task

这些具体特征,有明确的含义,可解释性很强,同时easy for ML

Raw Features

Raw features: often need human or machines to convert to concrete ones.

比如image pixels, speech signal等场景。

Abstract Features

Abstract: again need feature conversion/extraction/construction.

比如一些ID特征:

  • student ID in online tutoring system (KDDCup 2010)
  • advertisement ID in online ad system

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

相关文章

来自专栏码洞

人工不智能之sklearn聚类

线性回归是一种有监督算法,提供了输入数据x和参考目标值y,参考目标提供了一种纠错机制,是对预测结果y_的监督,如果y和y_相差过大,说明拟合的模型可能存在问题。...

641
来自专栏机器之心

教程 | 先理解Mask R-CNN的工作原理,然后构建颜色填充器应用

选自matterport 作者:Waleed Abdulla 机器之心编译 参与:刘晓坤 上年 11 月,matterport 开源了 Mask R-CNN 实...

3975
来自专栏机器之心

教程 | 使用Keras实现多输出分类:用单个模型同时执行两个独立分类任务

之前我们介绍了使用 Keras 和深度学习的多标签分类(multi-label classification),参阅 https://goo.gl/e8RXtV...

6432
来自专栏目标检测和深度学习

教程 | 先理解Mask R-CNN的工作原理,然后构建颜色填充器应用

选自matterport 作者:Waleed Abdulla 机器之心编译 参与:刘晓坤 上年 11 月,matterport 开源了 Mask R-CNN 实...

2375
来自专栏磐创AI技术团队的专栏

实用 | 分享一个决策树可视化工具

【磐创AI导读】:这篇文章希望跟大家分享一个可视化决策树或者随机森林的工具。这可以帮助我们更好的去理解或解释我们的模型。想要获取更多的机器学习、深度学习资源。欢...

2211
来自专栏机器之心

资源 | 一个Python特征选择工具,助力实现高效机器学习

项目地址:https://github.com/WillKoehrsen/feature-selector

1772
来自专栏杨熹的专栏

Ensemble Learners

Udacity Ensemble Learners ---- Boosting Algorithm 不需要绞尽脑汁去想很复杂的 Rules,只需要一些简单的 ...

3627
来自专栏数据派THU

教你在Python中用Scikit生成测试数据集(附代码、学习资料)

原文标题:How to Generate Test Datasets in Python with Scikit-learn 作者:Jason Brownlee...

5267
来自专栏挖数

简述【聚类算法】

所谓人以类聚,物以群分。人都喜欢跟自己像的人聚在一起,这些人或者样子长得比较像,或者身高比较像,或者性格比较像,或者有共同的爱好,也就是身上有某些特征是相似的。...

2936
来自专栏自学笔记

基于SVM的思想做CIFAR-10图像分类

回顾一下之前的SVM,找到一个间隔最大的函数,使得正负样本离该函数是最远的,是否最远不是看哪个点离函数最远,而是找到一个离函数最近的点看他是不是和该分割函数离的...

3633

扫码关注云+社区

领取腾讯云代金券