05 Training Versus Testing

Published: 08 Sep 2018 Category: ml-foundations

介绍Break Point等核心概念,见详细课件

两个核心问题

从Hoeffding不等式以及Union Bound得出:

学习可进一步被分成两个重要问题:

  1. Can we make sure that $E_{out}(g)$ is close enough to $E_{in}(g)$?
  2. Can we make sure $E_{in}(g)$ small enough?

Hoeffding inequality addresses the first question only.

复杂的模型更容易满足条件2,但同时会使1恶化;简单的模型更容易满足条件1,但无法保证条件2。简单地说,复杂模型能更好地拟合已知数据,但泛化能力较差;简单模型泛化能力强,反拟合已知数据能力弱。两个目标相互矛盾,所以:

The tradeoff in the complexity of $H$ is a majoy theme in learning theory.

有效假设

公式中$M$无限大,即坏事情发生的概率上界无限大,没有意义。无限大的来源是Union Bound,但条件太loose了,其实不同的假设重叠度非常高,可以:

Group similar hypotheses.

Dichotomy

对数据集合的二分。

表示假设空间对某数据集的所有二分。

ml-foundations-dichotomy

Growth Function

依赖特定数据集,所以取其上界,定义Growth function,摆脱对特定数据集的依赖,其定义在Hypothesis set $H$上

$m_H(N)$ is the maximum number of dichotomies that can be generated by $H$ on any $N$ points.

ml-foundations-growth-function

几个简单例子

ml-foundations-growth-function-example

Break Point

计算每个Hypothesis set的growth function不现实,再一次寻找其上界。

ml-foundations-break-point

不难得出$m_H(k) \lt 2^k$,这样只需要找到break point即可.例子

  • positive rays: break point at 2
  • positive intervals: break point at 3
  • convex sets: no break point
  • 2D perceptrons: break point at 4