跳到主要内容

東京大学 情報理工学系研究科 コンピュータ科学専攻 2015年8月実施 専門科目II 問題1

Author​

祭音Myyura (co-authored with GPT 5.6 SOL)

Description​

Consider the pattern recognition problem of classifying a dd-dimensional real vectorial pattern x∈Rdx\in\mathbb R^d into one of the two classes y=+1,−1y=+1,-1. For training a classifier, suppose that nn training samples

{(xi,yi)∣xi∈Rd, yi∈{+1,−1}, i=1,…,n}\{(x_i,y_i)\mid x_i\in\mathbb R^d,\ y_i\in\{+1,-1\},\ i=1,\ldots,n\}

are provided, where (xi,yi)(x_i,y_i) means that the pattern xix_i belongs to the class yiy_i.

Answer the following questions.

(1) Among the nn training samples, let n+n_+ and n−n_- be the numbers of patterns in the classes +1+1 and −1-1, respectively. Find the mean vector c+c_+ of the n+n_+ patterns in the class +1+1, and the mean vector c−c_- of the n−n_- patterns in the class −1-1.

(2) Consider the classifier that assigns a sample xx to the class +1+1 if ∥x−c+∥<∥x−c−∥\|x-c_+\|\lt\|x-c_-\|, and to the class −1-1 if ∥x−c+∥>∥x−c−∥\|x-c_+\|>\|x-c_-\|. Here ∥⋅∥\|\cdot\| denotes the Euclidean norm. Give an equation for the boundary between: the region to which the patterns classified into the class +1+1 belong; and the region to which the patterns classified into the class −1-1 belong.

For a parameter w∈Rdw\in\mathbb R^d, consider the linear classifier that assigns a sample xx to the class +1+1 if wTx>0w^\mathsf Tx>0, and to the class −1-1 if wTx<0w^\mathsf Tx\lt0. Here (⋅)T(\cdot)^\mathsf T denotes the transpose. Let us call the value yiwTxiy_iw^\mathsf Tx_i the margin for the ii-th training sample (xi,yi)(x_i,y_i). Then the condition for this linear classifier to correctly classify the pattern xix_i into the class yiy_i can be expressed, in terms of the margin, as yiwTxi>0y_iw^\mathsf Tx_i>0.

Answer the following questions.

(3) When the linear classifier shown above does not correctly classify the pattern xix_i into the class yiy_i, let us update the parameter ww by

wnew=w+yixi.w_{\mathrm{new}}=w+y_ix_i.

Prove that this parameter update does not decrease the margin for the ii-th training sample (xi,yi)(x_i,y_i).

(5) When the linear classifier shown above does not correctly classify the pattern xix_i into the class yiy_i, let us update the parameter ww by

wnew=arg⁡min⁡w′[∥w′−w∥2+(1−yiw′Txi)2].w_{\mathrm{new}} =\mathop{\arg\min}_{w'} \left[\|w'-w\|^2+(1-y_iw'^\mathsf Tx_i)^2\right].

Solve this optimization problem and obtain wneww_{\mathrm{new}} explicitly.

题目描述​

考虑把 dd 维实向量 x∈Rdx\in\mathbb R^d 分为 y=+1,−1y=+1,-1 两类的模式识别问题。给定训练样本

{(xi,yi)∣xi∈Rd,yi∈{+1,−1},i=1,…,n}.\{(x_i,y_i)\mid x_i\in\mathbb R^d, y_i\in\{+1,-1\}, i=1,\ldots,n\}.

(1)设两类样本数为 n+,n−n_+,n_-,求两类的均值向量 c+,c−c_+,c_-。

(2)最近均值分类器在 ∥x−c+∥<∥x−c−∥\|x-c_+\|\lt\|x-c_-\| 时判为 +1+1,在 ∥x−c+∥>∥x−c−∥\|x-c_+\|>\|x-c_-\| 时判为 −1-1。其中 ∥⋅∥\|\cdot\| 为欧几里得范数。求两类区域的边界方程。

线性分类器由 w∈Rdw\in\mathbb R^d 给出:wTx>0w^\mathsf Tx>0 判为 +1+1,wTx<0w^\mathsf Tx\lt0 判为 −1-1。其中 (⋅)T(\cdot)^\mathsf T 表示转置。样本 (xi,yi)(x_i,y_i) 的间隔为 yiwTxiy_iw^\mathsf Tx_i,正确分类的条件为 yiwTxi>0y_iw^\mathsf Tx_i>0。

(3)样本 xix_i 未被正确分类时,令 wnew=w+yixiw_{\rm new}=w+y_ix_i。证明该更新不会减小该样本的间隔。

(4)样本 xix_i 未被正确分类时,求下列优化问题的显式解:

wnew=arg⁡min⁡w′[∥w′−w∥2+(1−yiw′Txi)2].w_{\rm new}=\mathop{\arg\min}_{w'} \left[\|w'-w\|^2+(1-y_iw'^\mathsf Tx_i)^2\right].

Kai​

(1)​

两类均值要求 n+>0n_+>0 且 n−>0n_->0。

c+=1n+∑i:yi=+1xi,c−=1n−∑i:yi=−1xi.\boxed{ c_+=\frac1{n_+}\sum_{i:y_i=+1}x_i, \qquad c_-=\frac1{n_-}\sum_{i:y_i=-1}x_i.}

(2)​

两距离平方之差为

∥x−c−∥2−∥x−c+∥2=2(c+−c−)Tx+∥c−∥2−∥c+∥2.\|x-c_-\|^2-\|x-c_+\|^2 =2(c_+-c_-)^\mathsf Tx+\|c_-\|^2-\|c_+\|^2.

当 c+≠c−c_+\ne c_- 时,边界超平面为

2(c+−c−)Tx=∥c+∥2−∥c−∥2.\boxed{2(c_+-c_-)^\mathsf Tx=\|c_+\|^2-\|c_-\|^2.}

左边大于右边的一侧判为 +1+1,小于的一侧判为 −1-1。若 c+=c−c_+=c_-,则所有点到两均值的距离相等,题设的两个严格不等式均不成立,不能得到上述分类超平面。

(3)​

更新后的间隔为

yiwnewTxi=yiwTxi+yi2∥xi∥2=yiwTxi+∥xi∥2.y_iw_{\rm new}^\mathsf Tx_i =y_iw^\mathsf Tx_i+y_i^2\|x_i\|^2 =y_iw^\mathsf Tx_i+\|x_i\|^2.

因此间隔增加 ∥xi∥2≥0\|x_i\|^2\ge0,不会减小。

(4)​

目标函数严格凸。令其梯度为零,得到

w′−w=yixi(1−yiw′Txi).w'-w=y_ix_i(1-y_iw'^\mathsf Tx_i).

记 mi=yiwTxim_i=y_iw^\mathsf Tx_i、ri=∥xi∥2r_i=\|x_i\|^2。在上式左乘 yixiTy_ix_i^\mathsf T 并整理,可得

1−yiw′Txi=1−mi1+ri.1-y_iw'^\mathsf Tx_i=\frac{1-m_i}{1+r_i}.

所以唯一最优解为

wnew=w+1−yiwTxi1+∥xi∥2 yixi.\boxed{ w_{\rm new}=w+ \frac{1-y_iw^\mathsf Tx_i}{1+\|x_i\|^2}\,y_ix_i.}