Classification Short Version Hungyi Lee 1 Classification To

Classification (Short Version) Hung-yi Lee 李宏毅 1

Classification • To learn more …… https: //youtu. be/f. ZAZUYEe. IMg (in Mandarin) https: //youtu. be/h. SXFuyp. Luk. A (in Mandarin) 2

Classification as Regression? • Regression Model • Classification as regression? Model different? class 1 = class 1 2 = class 2 3 = class 3 similar? 3

Class as one-hot vector Class 3 Class 2 Class 1 or or + only output one value 1 + + 1 1 How to output multiple values? + 1 4

Class as one-hot vector or + + 1 1 Class 3 Class 2 Class 1 or 5

Regression feature label + + Classification feature + + label 0 or 1 Make all values between 0 and 1 Can have any value 6

Soft-max Softmax 0. 88 0. 12 ≈0 How about binary classification? 20 3 2. 7 1 0. 05 -3 logit 7

Loss of Classification softmax label Network Mean Square Error (MSE) Cross-entropy Minimizing cross-entropy is equivalent to maximizing likelihood. 8

http: //speech. ee. ntu. edu. tw/~tlkagk/courses/MLDS_2015_2/Lecture/Deep%20 More%20(v 2). ecm. mp 4/index. html softmax -10 ~ 10 large. Mean Square Error (MSE) loss -10 ~ 10 -1000 large loss Network Cross-entropy stuck! small loss Changing the loss function can change the difficulty of optimization. 9