D. Qiu的专栏

Model representationCost functionCost function intuition ICost function intuition IIGradient descentstart with some \theta_0, \theta_1keep changing \theta_0, \theta_1 to reduce

1. Learning with large datasetsIt's not who has the best algorithm that wins. It's who has the most data2. Stochastic gradient descentbatch gradient descentrepeat

1. Unsupervised learning introductionsupervised learning - training set: unpervised learning - training set: 2. K-means algorithmrandomly select K cluster centroidsrepeatfor every

1. Optimization Objectivelogistic regressionletsupport vector machine2. Large Margin Intuition3. The mathematics behind large margin classification (optional)

1. Motivation I: Data CompressionReduce data from 2D to 1D2. Motivation II: Data Visualization3. Principal Component Analysis problem formulationreduce from 2D to 1D: find a directio

1. Cost functionNeural Network (Classification)Binary classification1 output unitMulti-class classification (K classes)K output unitsCost function

1. Deciding what to try nextDebugging a learning algorithmSuppose you have implemented regularized linear regression to predict housing prices, when you test your hypothesis on a new set of

1. Prioritizing what to work on: Spam classification examplecollect lots of datadeveloped sophisticated features based on email routing informationdeveloped sophisticated features for message bo

1. Non-linear hypotheseswhy introduce non-linear hypotheses?high dimensional datanon-linear hypotheses2. Neurons and the brainNeural networksorigins: algorithms that try to m

1. Classification2. Hypothesis RepresentationLogistic Regression Modelthe above function is called sigmoid function or logistic functionInterpretation of Hypothesis outputestimated probability tha

Multiple featuresfor convenience of notation, define , the new hypothesis isGradient descent for multiple variablesnew algorithm isGradient descent in practive I: Feature S

1、线性回归、logistic回归和一般回归线性回归函数（如果）损失函数（loss function）或者错误函数（error function）梯度下降法梯度下降法最大的问题是解有可能是局部极小值，这与初始点的选取有关步骤首先对赋值，这个值可以是随机的，也可以让是一个全零的向量改变的值，使得沿梯度下降的的方向减少由于求得的是极小值，因此