PAC
Probably Approximately Correct
Definition
$\mathcal{X}$:input space (the set of all possible examples or instances)
$\mathcal{Y}$:output space (all possible labels or target values)
A concept $c \colon \mathcal{X} \to \mathcal{Y}$ is a mapping from $\mathcal{X}$ to $\mathcal{Y}$.
对于可以学到的 mapping 称作 hypothesis
它们属于对应集合,concept $c\in \mathcal{C}$ 以及 hypothesis $h\in \mathcal{H}$。
consistent vs computational cost trade-off:$\mathcal{H}$ 越大,$\mathcal{C,H}$ 交集越大 (consistent),但是对应计算复杂度更大 (computational cost)
同时 $\mathcal{H}$ 过大/小也会出现 overfitting 或者 underfitting 的问题。
Definition (PAC-learning): A concept class $\mathcal{C}$ is PAC-learnable if there exists an algorithm $\mathcal{A}$ such that for all $\varepsilon > 0$, $\delta > 0$, for all distributions $\mathcal{D}$, and for all $c \in \mathcal{C}$:
- $n$:表示 Sample $x$ 的维度/编码复杂度;
- $\text{size}(c)$:表示描述 concept $c$ 或者 hypothesis 的表示大小/复杂度;
例如在二分类问题中,对于 $n$ 个元素而言想要分辨出谁是 1 谁是 0,需要 $\mathbb{F}_2^n$ 下的一个向量描述,共 $n$ bit 所以 $\text{size}(c)=n$。
- $R(h_S) \le \varepsilon$:近似正确(误差率不超过容忍度 $\varepsilon$)Approximately Correct ;
- $\ge 1 - \delta$:以至少 $1 - \delta$ 的高置信度成立 Probably.
也就是说,在样本数 $m$ 足够大时,在 $\mathcal{D}$ 中采样得到 $\mathcal{S}$,其 Generalization Error 是 Probably Approximately Correct 的。
更进一步地, if $\mathcal{A}$ runs in $\text{poly}\left(\frac{1}{\varepsilon}, \frac{1}{\delta}, n, \text{size}(c)\right)$,则称其 efficiently PAC-learnable .
注意这里是对一个 concept 集合 $\mathcal{C}$ 而言的,不是对特定的一种情况 concept $c$。
Unbiased Estimator
设在某应用场景下(例如观测赤道地区的温度指标),观测样本 $X_1, \dots, X_n$ 为独立同分布(i.i.d.)的随机变量:
- $\mathbb{E}[X_i]=\mu$
- $\text{Var}[X_i] = \sigma^2$
定义令 Empirical Mean 作为估计量(预测模型): $\bar{X} = \frac{1}{n} \sum_{i=1}^n X_i$,则模型的期望与方差分别为:
- $\mathbb{E}[\bar{X}] = \mathbb{E}\left[ \frac{1}{n} \sum_{i=1}^n X_i \right] = \frac{1}{n} \sum_{i=1}^n \mathbb{E}[X_i] = \frac{1}{n} \cdot (n \mu) = \mu$
- $\text{Var}[\bar{X}] = \text{Var}\left( \frac{1}{n} \sum_{i=1}^n X_i \right) = \frac{1}{n^2} \sum_{i=1}^n \text{Var}[X_i] = \frac{1}{n^2} \cdot (n \sigma^2) = \frac{\sigma^2}{n}$
Concentration(测度集中):
Chebyshev’s Inequality: 对于任意随机变量 $X$ 及其偏离阈值 $t > 0$ 有: $\Pr(|X - \mathbb{E}[X]| \ge t) \le \frac{\text{Var}[X]}{t^2}$
令 $t = \varepsilon \mu$,$\Pr(|\bar{X} - \mu| \ge \varepsilon \mu) \le \frac{\text{Var}[\bar{X}]}{(\varepsilon \mu)^2} = \frac{\frac{\sigma^2}{n}}{\varepsilon^2 \mu^2} = \frac{\sigma^2}{\varepsilon^2 \mu^2 n}$。
PAC框架中,需要单次估计至少 $1 - \delta$ 的置信度,则
其中 sample 度量维度为 $1$,因为只有一维标量 label (or value) 即预测温度;由于直接给出预测值 hypothesis size 视作常数。
ERM 任务,使用最小二乘损失,其解为 sample mean,model size 视作常数。
Example 2.4 (Learning axis-aligned rectangles)
下面这个解释是 ok 的,来自该视频的评论区。

其中《注意到》简而言之是在说明,$R_S$ 四条边都分别落在 $r_1,r_2,r_3,r_4$ 中 $\Longrightarrow$ 误差 $\leq \epsilon$,则讨论误差 $>\epsilon$ 的概率仅需考虑其逆否命题。
假设二分类数据集,模型是记忆所有的正类样本。遇到相同的就为正,否则为负。
Generalization Bounds
Theorem 2.5 (Learning bound — finite $\mathcal{H}$, consistent case)
Theorem 2.5 (Learning bound — finite $\mathcal{H}$, consistent case),考虑 hypothesis set $\mathcal{H}$ finite 且 consistent 的情况下,let $\mathcal{A}$ be the algorithm that for any concept $c\in \mathcal{H}$ and I.I.D. Sample $S$,$\mathcal{A}$ 给出的 $h_S$ 满足 Training Error $\widehat{R}_S(h_S) = 0$,则 for any $\epsilon,\delta>0$,概率 $\mathbb{P}_{S \sim \mathcal{D}^m} [R(h_S) \le \epsilon] \ge 1 - \delta$ holds if $m \ge \frac{1}{\epsilon} \left( \log |\mathcal{H}| + \log \frac{1}{\delta} \right)$。
将其改写为等价的 Generalization bound 的形式(由于任意满足条件的 $\epsilon$ 均成立,故可取 $\epsilon=\frac{1}{m}\left( \log |\mathcal{H}| + \log \frac{1}{\delta} \right)$,此时原命题等价于
Proof: 考虑问题的反面,如果学到的 $h_S$ 满足 $R(h_S)>\epsilon$,说明 $h_S$ 在分布 $D$ 中有 $>\epsilon$ 的概率出错。而 $S$ 是 I.I.D. 选取的,则 $Pr\left(\widehat{R}_S(h_S) = \frac{1}{m}\sum\limits_{i=1}^m\mathbb{1}_{h(x_i)\neq y_i}=0\right)\leq (1-\epsilon)^m$,即每次抽样都抽到表现正确的 example。令 $\mathcal{H}_\epsilon = \{h \in \mathcal{H}: R(h) > \epsilon\}$,则失败即选取到 $\mathcal{H}_{\epsilon}$ 内满足 $\widehat{R}_S(h) = 0$ 的 $h$,那么计算存在这样一个 $h$ 的概率。
令 $|\mathcal{H}|e^{-m\epsilon}\leq \delta$ (失败率) 则得到 $m\geq \frac{1}{\epsilon}(\log H+\log\frac{1}{\delta})$,PAC-learning。
所以,Under finite $H$, consistent algorithm $A$ is PAC-learnable algorithm.
回顾一下,如果 Algorithm $\mathcal{A}$ 总能找到 consistent hypothesis, 并且 sample 足够大时,它在泛化误差上的表现不会太差,要不然它是以很低的概率才能恰好命中自己会的那些 training data。
[Example 2.6] Conjunction of Boolean literals
要预测一个输入在合取范式下的结果 (positive, or negative),给定很多 examples 有 positive 和 negative 的。执行算法:先假设所有文字和其反面都在合取范式中,每次得到一个 positive example 时将其违反的文字去除。例如 $(1,1,0,0)$ 会将 $\bar{x_1},\bar{x_2},x_3,x_4$ 去除。那么得到的合取范式肯定会对 positive example 预测正确,由于可能去除本应该去除的文字,所以该合取范式会比真正的合取范式更严格,那么对 negative example 也可以预测正确。
现在我们来看 hypothesis set $\mathcal{H}$,对每个文字存在三种情况:$x,\bar x,\empty$,分别要求 positive, negative, unknown,所以 $|\mathcal{H}|=|\mathcal{C}|=3^n$,根据定理,有 $m \ge \frac{1}{\epsilon} \left( (\log 3)n + \log \frac{1}{\delta} \right)$,PAC-learnable,同时也是 efficient PAC-learnable。
Learning bound — finite $\mathcal{H}$, inconsistent case
Hoeffding’s inequality: 设 $X_1, X_2, \dots, X_n$ 为相互独立的随机变量,且每个 $X_i$ 几乎必然取值于有界闭区间 $[a_i, b_i]$。其均值 $\bar{X} = \frac{1}{n} \sum_{i=1}^n X_i$,对于任意常数 $t > 0$,单侧偏差界 $P(\bar{X} - \mathbb{E}[\bar{X}] \ge t) \le \exp\left( - \frac{2 n^2 t^2}{\sum_{i=1}^n (b_i - a_i)^2} \right)$,其几何意义是测度几乎集中在均值附近。
在 inconsistent case 中 $\widehat{R}_S(h)\neq 0,\mathbb{E}\left(\widehat{R}_S(h)\right)=R(h)$,那么就可以用 Heffding’s inequality 去 bound 住 $R(h)$,对于 $\widehat{R}_S(h)=\sum\frac{1}{m}\mathbb{1}_{h_S(x_i)\neq y_i}$ 则 $\sum(b_i-a_i)^2=\frac{1}{m}$ 所以有 $\underset{S \sim \mathcal{D}^m}{\Pr} \left( \left\vert{} \widehat{R}_S(h) - R(h) \right\vert{} \ge \varepsilon \right) \le 2\exp(-2m\varepsilon^2)$ 令其 $=\delta$,所以化简得到
$\Rightarrow$ single hypothesis: $h \colon \mathcal{X} \to \{0, 1\}$, $\forall \delta > 0$, w.p. $\ge 1 - \delta$:
Theorem 2.13 (Learning bound — finite H, inconsistent case)
Theorem 2.13 (Learning bound — finite $\mathcal{H}$, inconsistent case) Let $\mathcal{H}$ be a finite hypothesis set. Then, for any $\delta > 0$, with probability at least $1 - \delta$, the following inequality holds:
Proof (Obviously):
利用 Union Bound,再令最后 bound 的概率为 $\delta$,最后解得 $\epsilon = \sqrt{\frac{\log \vert{}\mathcal{H}\vert{} + \log \frac{2}{\delta}}{2m}}$,那么
对 hypothesis class 内所有 hypothesis $h \in \mathcal{H}$ 同时成立 w.p. $\geq 1-\delta$:
Agnostic PAC-learning
当输出是随机分布的。
Definition 2.14 (Agnostic PAC-learning) Let $\mathcal{H}$ be a hypothesis set. $\mathcal{A}$ is an agnostic PAC-learning algorithm if there exists a polynomial function $\operatorname{poly}(\cdot, \cdot, \cdot, \cdot)$ such that for any $\epsilon > 0$ and $\delta > 0$, for all distributions $\mathcal{D}$ over $\mathcal{X} \times \mathcal{Y}$, the following holds for any sample size $m \ge \operatorname{poly}(1/\epsilon, 1/\delta, n, \operatorname{size}(c))$:
If $\mathcal{A}$ further runs in $\operatorname{poly}(1/\epsilon, 1/\delta, n)$, then it is said to be an efficient agnostic PAC-learning algorithm.
Rademacher Complexity
Definition
The family of loss functions associated with $\mathcal{H}$: $\mathcal{G} = \{g \colon (x, y) \mapsto L(h(x), y) \colon h \in \mathcal{H}\}$,即对于所有 hypothesis $h$,它对应的 loss function 构成的集合。
Definition 3.1 (Empirical Rademacher complexity) Let $\mathcal{G}$ be a family of functions mapping from $\mathcal{Z}$ to $[a, b]$ and $S = (z_1, \dots, z_m)$ a fixed sample of size $m$ with elements in $\mathcal{Z}$. Then, the empirical Rademacher complexity of $\mathcal{G}$ with respect to the sample $S$ is defined as:
其中 $\sigma$ 是一个由 $\{-1,+1\}$ 各 w.p. $\frac{1}{2}$ 构成的向量,为随机噪声。$\widehat{\mathfrak{R}}_S(\mathcal{G}) = \underset{\boldsymbol{\sigma}}{\mathbb{E}}\left[\sup_{g \in \mathcal{G}} \frac{\boldsymbol{\sigma} \cdot \mathbf{g}_S}{m}\right]$,其中内积是衡量相似度的指标,所以此处反映了 loss functions family $\mathcal{G}$ 对随机噪声的拟合能力 / 表达能力。
Definition 3.2 (Rademacher complexity)
Theorem 1
对 i.i.d. Sample $\mathcal{S}$ of size $m$, w.p. $\geq 1-\delta$, $\forall g\in \mathcal{G}$,存在泛化界
Proof: 定义 $g$ 在 $S$ 上的 empirical average $\widehat{\mathbb{E}}_S[g] = \frac{1}{m} \sum_{i=1}^m g(z_i)$. 对 Sample $S$ 定义
由于 $\Phi(S) = \sup_{g \in \mathcal{G}} \left(\mathbb{E}[g] - \widehat{\mathbb{E}}_S[g]\right)\geq \mathbb{E}[g] - \widehat{\mathbb{E}}_S[g],\forall g\in \mathcal{G}$,对 $(3.3)$ 仅需证明 $\Phi(S)\leq 2\mathfrak{R}_m(\mathcal{G}) + \sqrt{\frac{\log \frac{1}{\delta}}{2m}}$。
McDiarmid 不等式
对于 $m$ 个独立随机变量 $X_i \in \mathcal{X}$,若函数 $f$ 满足 Lipschitz condition 即 $\vert{}f(X_1, \cdots, X_{i-1}, X_i, X_{i+1}, \cdots, X_m) - f(X_1, \cdots, X_{i-1}, X_i’, X_{i+1}, \cdots, x_m)\vert{} \le c_i$,则对于任意 $\varepsilon > 0$,有:
当 $f$ 为 $\sum$ 时就是 Hoeffding 不等式。
Reference/Bounded_Difference_Method)
考察 $\Phi(S)$ 是否满足 Lipschitz 条件,它是在随机变量 $S=(S_1,S_2,\ldots,S_m)$,那么 $\Phi(S’)-\Phi(S)=\sup\left(\widehat{\mathbb{E}}_S[g]-\widehat{\mathbb{E}}_S’[g]\right)=\sup\left(\frac{1}{m}\left(g(z_i)-g(z’_i)\right)\right)=\frac{1}{m}$,此处不妨 bound 住损失函数 $\mathcal{L}: \mathcal{Y}\times \mathcal{Y}\to \mathcal{R}_{[0,1]}$(因为 $\mathcal{R}$ 与 $\mathcal{R}_{[0,1]}$ 存在双射例如 $\arctan$)。
所以 $\Phi(S)$ 是 $\frac{1}{m}$ Lipschitz 的,故 $\Pr\left(\Phi(S)\leq \mathbb{E}\left[\Phi (S)\right]+t\right) \le \exp\left(-2m{t^2}\right)=\delta$,解得 $t=\sqrt{\frac{\log \frac{1}{\delta}}{2m}}$,w.p. $\geq 1-\delta$。
则 $\Phi(S)\leq \mathbb{E}\left[\Phi (S)\right]+\sqrt{\frac{\log \frac{1}{\delta}}{2m}}$ w.p. $\geq 1-\delta$,现仅需证 $\mathbb{E}\left[\Phi (S)\right]\leq 2\mathfrak{R}_m(\mathcal{G})$
$(3.8)$ 到 $(3.9)$ 是由于 $\sup$ 存在 sub-additivity 次可加性,即 $\sup f+\sup g\geq \sup(f+g)$,把内部的期望拆出来即可。其中 $(3.10)$ 到 $(3.11)$ 中引入了噪声,因为 $S$ 和 $S’$ 都是 i.i.d. 的,所以任意交换 $z$ 和 $z’$ 的位置不会影响期望。 $(3.12)$ 依然是利用 sub-additivity。自此,证明了 $\Phi(S)\leq 2\mathfrak{R}_m(\mathcal{G})+\sqrt{\frac{\log \frac{1}{\delta}}{m}}$ w.p. $\geq 1-\delta$,则 $(3.3)$ 自然成立。
对于 $(3.4)$,对 $\mathfrak{R}_m(\mathcal{G}) = \mathbb{E}[\widehat{\mathfrak{R}}_S(\mathcal{G})]$ 运用 McDiarmid inequality 得到 w.p. $\geq 1-\delta$,$\mathfrak{R}_m(\mathcal{G})\leq \widehat{\mathfrak{R}}_S(\mathcal{G})+\sqrt{\frac{\log\frac{1}{\delta}}{2m}}$。为了统一 probability,在 $(3.3)$ 和此处均取 w.p. $\geq 1-\frac{\delta}{2}$,那么均满足的概率 $\geq 1-\delta$。,此时联合 $\mathbb{E}[g(z)] \le \frac{1}{m} \sum_{i=1}^m g(z_i) + 2\mathfrak{R}_m(\mathcal{G}) + \sqrt{\frac{\log \frac{2}{\delta}}{2m}}$ 和 $\mathfrak{R}_m(\mathcal{G})\leq \widehat{\mathfrak{R}}_S(\mathcal{G})+\sqrt{\frac{\log\frac{2}{\delta}}{2m}}$,就得到 $(3.4)$。
Theorem 3.3 给予我们一个通过 Rademacher complexity 用 empirical error 给 generalization error 定界的一个方法。
Theorem 2
Lemma 3.4 Let $\mathcal{H}$ be a family of functions taking values in $\{-1, +1\}$ and let $\mathcal{G}$ be the family of loss functions associated to $\mathcal{H}$ for the zero-one loss: $\mathcal{G} = \{(x, y) \mapsto 1_{h(x) \neq y} : h \in \mathcal{H}\}$. 即 $\{-1,+1\}$ 的损失函数与对应 01 损失函数集合,empirical error 呈两倍关系
For any sample $S = ((x_1, y_1), \dots, (x_m, y_m))$ of elements in $\mathcal{X} \times \{-1, +1\}$, let $S_{\mathcal{X}}$ denote its projection over $\mathcal{X}$: $S_{\mathcal{X}} = (x_1, \dots, x_m)$. Then, the following relation holds between the empirical Rademacher complexities of $\mathcal{G}$ and $\mathcal{H}$:
Proof: 对于 $y, h(x) \in \{-1, +1\}$,01 损失可以表示为 $1_{h(x) \neq y} = \frac{1 - y \cdot h(x)}{2}$ 代回原式得证。
Theorem 3.5 (Rademacher complexity bounds – binary classification) $\mathcal{H}$ 为假设空间,$h\in \mathcal{H}$ 满足 $h(x)\in \{-1,+1\},\forall \delta>0$:
由 Theorem 3.3 和 Lemma 3.4 即证。为了让 bound 住的界更紧,即 minimize $\underset{\boldsymbol{\sigma}}{\mathbb{E}}\left[\inf\limits_{g \in \mathcal{G}} \frac{1}{m} \sum\limits_{i=1}^m \sigma_i g(z_i)\right]$,这个问题叫作 Empirical Risk Minimization (ERM),对 classification problem with a 0-1 loss function 是 NP-Hard 问题。