Lecture 4 - AsyDynamics/CS231n GitHub Wiki

Back propagation

• Local gradient $/times$ upstream gradient
• Combine some of the operation as pattern, e.g. the sigmoid function
• add gate - gradient distributor
• max gate - gradient router
• mul gate - gradient switcher
• Jacobian matrix
• API forward backward

Neural netword

• before: linear score function
• now, with multiple layer
• activation function, many types