The weights of a neural network cannot be calculated using an analytical method. Instead, the weights must be discovered via an empirical optimization procedure called stochastic gradient descent.
Stochastic gradient descent is an optimization algorithm that estimates the error gradient for the current state of the model using examples from the training dataset, and then updates the weights of the model using backpropagation.
Learning rate is a hyper-parameter that controls the weights of our neural network with respect to the loss gradient. It defines how quickly the neural network updates the concepts it has learned.
A desirable learning rate is low enough that the network converges to something useful, but high enough that it can be trained in a reasonable amount of time.
Smaller learning rates require more training epochs (requires more time to train) due to the smaller changes made to the weights in each update, whereas larger learning rates result in rapid changes and require fewer training epochs. However, larger learning rates often result in a sub-optimal final set of weights.
Adaptive learning rates allow the training algorithm to monitor the performance of the model and automatically adjust the learning rate for the best performance.
The most basic model of this decreases the learning rate once the performance of the model reaches a plateau. The model does so by decreasing the learning rate by a factor of two or an order of magnitude. However, the learning rate can be increased again if the performance doesn’t improve.
Neural networks with adaptive learning rates usually outperform ones with fixed learning rates.
View all Courses