【李宏毅机器学习课件】20.SVM (v5).pdf

Support Vector Machine Outline Hinge Loss Support Vector Machine (SVM) Kernel Method 1 2 3 肀 肀 肀 鈥︹€ Binary Classification 肀︵窚1 肀︵窚2 肀︵窚3 饊仜 肀︵窚 = +1, 鈭? 鈥 Step 1: Function set (Model) 肀 肀 0 Output = +1 肀 肀 = 肀 肀 0 Output = -1 鈥 Step 2: Loss function: The number of times g 戆 肀 = 喾嶍豢 肀 肀ヰ€仜 鈮犿宝喾滒€仜 get incorrect results on 饊仜 饊仜 饊仜 training data. 肀 肀 肀 ,肀︵窚 鈥 Step 3: Training by gradient descent is difficult Gradient descent is possible if g 鈭 and 砘 鈭 is differentiable Step 2: Loss function 肀 肀 = 肀 肀 0 Output = +1 肀 肀 0 Output = -1 Ideal loss: Approximation: 饊仜 饊仜 饊仜 饊仜 戆 肀 =喾嶍豢 肀 肀 鈮犿宝喾 戆 肀 =喾 肀 肀 肀 ,肀︵窚 饊仜 饊仜 饊仜 饊仜