-
Notifications
You must be signed in to change notification settings - Fork 30
15. Training Neural Networks
Topic: Training Neural Networks
Course: GMLC
Date: 31 March 2019
Professor: Not specified
-
https://developers.google.com/machine-learning/crash-course/training-neural-networks/video-lecture
-
https://developers.google.com/machine-learning/crash-course/training-neural-networks/best-practices
-
https://developers.google.com/machine-learning/crash-course/backprop-scroll/
-
Backpropagation
-
Algorithm used to train Neural Networks
-
Performs gradient descent
-
Steps:
-
Calculates outputs of all nodes in a forward pass
-
Calculates error rate using derivatives of the error with each parameter
-
-
-
Failure cases
-
Vanishing gradients
-
Gradients vanish towards 0 and train very slowly or not at all
-
Rectified Linear Unit activation function can prevent the problem
-
-
Exploding gradients
-
Large weights can cause the lower layers of the network to involve too many products, thus they get too large to converge
-
Lowering the learning rate or batch normalization prevents the problem
-
-
Dead ReLU units
-
Weighed sum of ReLU activation function can fall below 0, thus not contributing to neural network at all
-
Lowering the learning rate prevents dead ReLU units
-
-
-
Dropout regularization
- Works by randomly dropping out unit activations
-
Explain how backpropagation works
-
Be aware of NN failure cases
- Neural networks are trained using backpropagation algorithm which sums the outputs of each node and runs them through derivatives for error calculation & through activation functions for learning