Please enable JavaScript.
Coggle requires JavaScript to display documents.
Logistic Regression (:sweat_drops:limitations (:maple_leaf:坐标转换(特征提取) (…
Logistic Regression
:moneybag:vs. Linear Regression
:one:differences
:two:common
:explode:2种regression的计算方式和采用的损失函数都不同,但最后导致的更新方式相同
:moneybag:update
:warning:名义上是regression,实际上是classification
loss function
:explode:cross entropy
:poultry_leg:为什么是cross entropy?
:explode:极大似然估计+0,1条件筛选
:one:定义极大似然估计函数
:bread:这里的\( 1 - f_{w,b}(x^3) \)是因为\( x^3 \)是属于C2的,所以利用0,1性质知道用1减去就是C1的概率
:two:用ln把乘法转换成加法,并且用\( \hat{y}^i \)来条件筛选
:three:得到最终的损失函数(极大似然估计的相反)
:poultry_leg:为什么不使用Square Error?
:explode:使用SE得到的更新方式会在距离真实值远的地方更新也可能慢
:cherry_blossom:theory
:egg:对w和b的极大似然估计+标签只有0和1
:globe_with_meridians:probabilistic generative model
:reminder_ribbon:PGM(generative model) & logistic regression(discriminative model)
:notes:Generally speaking, discriminative model performs better than generative model
:palm_tree:
:icecream:example
:bread:使用朴素贝叶斯, 假定2个feature独立,那么即使直觉告诉我们属于C1,但是由于C2样本量大,而且2个feature独立,导致认为C2概率大
:warning:当然也可能确实属于C2
:link:
details
:sweat_drops:limitations
:explode:logistic regression不是总能将若干点分成2类
:icecream:异或
:zap:sigmoid以0.5区分=>\(wx+b\)以0区分 => 线性区分
:maple_leaf:坐标转换(特征提取)
:one:坐标转换矩阵
:explode:级联logistic regression
:arrow_right:Neural Network
:warning:如果是按照logistic regression思维的话,activation应该是sigmoid,但是现在更多的是其他activation function
:icecream:example
:two:某个点的新坐标的某个维度的值=原来这个点到某个点的位置
