图1-16 分类年龄-收入数据Classifying the age and income data
分类和回归都是有监督学习的一部分。
Both classification and regression areparts of supervised learning.
因此，它们的训练数据同样是{输入，正确输出}的形式。
Therefore, their training data is equallyin the form of {input, correct output}.
二者唯一的区别是正确输出的结果，分类输出数据的类别，而回归输出的是估计值。
The only difference is the type of correctoutputs—classification employs classes, while the regression requires values.
总之，当需要模型判断输入数据属于哪个分组时，分析过程就变成分类；当模型估计数据的趋势时，分析过程就变成了回归。
In summary, analysis can becomeclassification when it needs a model to judge which group the input databelongs to and regression when the model estimates the trend of the data.
作为参考，无监督学习的代表性应用之一是聚类。
Just for reference, one of therepresentative applications of unsupervised learning is clustering.
聚类研究了个体数据的特征，并对相关数据进行分类。
It investigates the characteristics of theindividual data and categorizes the related data.
聚类和分类很容易混淆，因为它们的结果是相似的。
It is very easy to confuse clustering andclassification, as their results are similar.
虽然它们产生相似的输出，但它们是两种完全不同的方法。
Although they yield similar outputs, theyare two completely different approaches.
我们必须记住，聚类和分类是不同的术语。
We have to keep in mind that clustering andclassification are distinct terms.
当你遇到术语聚类时，只需提醒自己，它专注于无监督的学习。
When you encounter the term clustering,just remind yourself that it focuses on unsupervised learning.
小结
Summary
让我们简要回顾一下本章所涵盖的内容：
Let’s briefly recap what we covered in thischapter:
人工智能、机器学习和深度学习是不同的。
Artificial Intelligence, Machine Learning,and Deep Learning are distinct.
但它们之间存在以下联系：“深度学习是一种机器学习，而机器学习是一种人工智能”。
But they are related to each other in thefollowing way: “Deep Learning is a kind of Machine Learning, and MachineLearning is a kind of Artificial Intelligence”.
机器学习是一种从训练数据中导出模型的归纳方法。
Machine Learning is an inductive approachthat derives a model from the training data.
对图像识别、语音识别、自然语言处理等都有一定的参考价值。
It is useful for image recognition, speechrecognition, and natural language processing etc.
机器学习的成功很大程度上依赖于如何实现泛化过程。
The success of Machine Learning heavilyrelies on how well the generalization process is implemented.
为了防止由于训练数据与实际输入数据之间的差异导致的性能下降，我们需要足够数量的无偏训练数据。
In order to prevent performance degradationdue to the differences between the training data and actual input data, we needa sufficient amount of unbiased training data.
当模型过度适应训练数据时将发生过度拟合，即它对实际输入数据产生不良的性能，而对训练数据的性能是极其优异的。
Overfitting occurs when the model has beenoverly customized to the training data that it yields poor performance for theactual input data, while its performance for the training data is excellent.
过度拟合是降低泛化性能的主要因素之一。
Overfitting is one of the primary factorsthat reduces the generalization performance.
正则化和验证是解决过度拟合问题的典型方法。
Regularization and validation are thetypical approaches used to solve the overfitting problem.
正则化是一种能产生最简单模型的数值方法。
Regularization is a numerical method thatyields the simplest model as possible.
与此相反，验证试图在训练过程中发现过度拟合的迹象，并采取措施进行防止。
In contrast, validation tries to detectsigns of overfitting during training and takes action to prevent it.
——本文译自Phil Kim所著的《Matlab Deep Learning》
更多精彩文章请关注微信号：
