Hexo

低维空间内，线性不可分时，可以采用非线性。这样等同于映射到高维空间，然后采用线性分类器分类。

比如二类分界面是 y=3x^2+2x+6,我们可采用把原始x映射到（1，x，x^2）空间中，进行线性分类。

核函数可以简单理解成 一种相似性度量
大部分核函数可以归结为两种，内积和1/d
比如矩阵分解中，两个向量的内积来作为相似性度量。
高斯核函数是， 1/exp(d)

也可以用KL距离吧

IsanSVMassimpleassayingit'sadiscriminativeclassifierthatsimplyoptimizesthehingeloss?

SVMissimplya**linearclassifier,optimizinghingelosswithL2regularization**.

Orisitmorecomplexthanthat?

No,itisjustthat,howevertherearedifferentwaysoflookingatthismodelleadingtocomplex,interestingconclusions.Inparticular,thisspecificchoiceoflossfunctionleadstoextremelyefficientkernelization,whichisnottrueforlogloss(logisticregression)normse(linearregression).Furthermoreyoucanshowveryimportanttheoreticalproperties,suchasthoserelatedtoVapnik-Chervonenkisdimensionreductionleadingtosmallerchanceofoverfitting.

支持向量有什么作用？

Supportvectorsaresimplysamplesplacednearthedecisionboundary(loselyspeaking).Forlinearcaseitdoesnotchangemuch,butasmostofthepowerofSVMliesin**itskernelization**-thereSVsareextremelyimportant.Onceyouintroducekernel,duetohingeloss,SVMsolutioncanbeobtained**efficiently**,andsupportvectorsaretheonlysamplesrememberedfromthetrainingset,thusbuildinganon-lineardecisionboundarywiththesubsetofthetrainingdata.

为什么CNN通常采用logloss(softmax)，为什么很少采用hingeloss?CNN+hinge-loss能叫deepSVM吗?

https://www.reddit.com/[object Object]comments/40680m/why_isnt_cnns_with_hinge_loss_popular_can_we_call/

hingeloss和ReLU很像，有什么关系吗？

参考

• 很经典，很赞，常见SVM的疑问 https://stackoverflow.com/questions/34325759/whats-the-relationship-between-an-svm-and-hinge-loss