参考：
【1】机器学习理论表明，机器学习算法能从有限个训练集样本上得到较好的泛化
【1】Machine learning theory shows that machine learning algorithm can generalize well from finite training set samples
- limited有限的
- infinite无限的
【2】这似乎违背了一些基本的逻辑准则
【2】This seems to violate some basic rules of logic
- go against 违背
【3】如果我们要从逻辑上推理出一个规则以描述集合中所有元素，必须知道该集合中所有元素的信息
【3】If we want to logically reason out a rule that describes all the elements of a set, we must know all the elements of that set
- inference:（名词）推理
【4】在一定程度上，机器学习通过使用概率规则，而不是完全确定性法则来避免这个问题
【4】To some extent, machine learning avoids this problem by using probabilistic rules rather than fully deterministic rules
- probability 概率
- certain 确定
- determinate 确定
【5】不幸的是，即使这样也不能解决整个问题。机器学习中的“没有免费午餐”定理表明，每一个分类算法对未实现观测的点进行分类的错误率，在所有可能的数据生成分布上的均值是相等的。
【5】Unfortunately, even that won't solve the whole problem. The "no free lunch" theorem in machine learning shows that the error rate of each classification algorithm in classifying unrealized observation points is equal on the mean value of all possible data generation distributions.
【6】换言之，在某种意义上，没有一个机器学习算法总是比其他算法要好
【6】In other words, there is no machine learning algorithm which is always better than others in some sense
【7】我们所能设想的最复杂精致的算法，与一个仅仅将所有数据归为一类的算法，在所有可能的数据生成分布上有着相同的平均性能。
【7】The most sophisticated algorithm we can design has the same average performance across all possible data generation distributions as an algorithm that simply groups all data into one category.
【8】这意味着，机器学习的研究目标不是找一个通用的或绝对最好的学习算法
【8】This means that the goal of machine learning research is not to find a universal or absolute best learning algorithm
【9】反之，我们的目标是理解什么样的机器学习算法在我们关心的数据生成分布上表现最好
【9】Instead, our goal is to understand which machine learning algorithms perform best on the data generation distribution that we care about
