||
这一讲主要是分析上一讲中联合错误界中常数$M$的分析。在上一讲中,我们得知
If $|H|=M$ finite, $N$ large enouph, for wahtever $g$ picked by $A$, $E_{out}(g)\approx E_{in}(g)$;
If $A$ finds one $g$ with $E_{in}(g)\approx 0$, PAC guarantee for $E_{out}(g)\approx 0$.
$\Rightarrow $ Learning possible.
learning split to two central questions:
can we make sure that $E_{out}(g)$ is close enough to $E_{in}(g)$?
can we make $E_{in}(g)$ small enough?
2.我们已知了错误上界,现在的目的在于如何限定$M$使得错误界更加紧凑。
3. 利用集合的联合导出的上述错误界,其实是对学习错误的过高估计,因为:
4.Growth Function:remove dependence by taking max of all possible $(x_1; x_2;\cdots; x_N)$
$m_H(N) = \max\limits_{x_1;x_2;\cdots;x_N\in \mathcal{X}} |H(x_1; x_2; \cdots ; x_N)|$
其中,$|H(x_1; x_2; \cdots ; x_N)|$是这N个样本产生的假设的个数。
$m_H(N)$ finite and upper-bounded by 2N.
下图是几个学习算法的成长函数(Growth Function)值:
由此,我们可以猜测
5.Break Point of $\mathcal{H}$
if no $k$ inputs can be shattered by $H$, call $k$ a break point for $H$:
$m_H(k) < 2^k$;
$k + 1, k + 2, k + 3, \cdots$ also break points!
we will study minimum break point $k$.
Break point 的几个例子:
6. 最后给出假设,引出下一讲
Archiver|手机版|科学网 ( 京ICP备07017567号-12 )
GMT+8, 2024-11-25 02:50
Powered by ScienceNet.cn
Copyright © 2007- 中国科学报社