AdaDelta

Optimizer that implements the Adadelta algorithm.

Adadelta optimization is a stochastic gradient descent method that is based on adaptive learning rate per dimension to address two drawbacks:

• the continual decay of learning rates throughout training
• the need for a manually selected global learning rate

Adadelta is a more robust extension of Adagrad that adapts learning rates based on a moving window of gradient updates, instead of accumulating all past gradients. This way, Adadelta continues learning even when many updates have been done. Compared to Adagrad, in the original version of Adadelta you don't have to set an initial learning rate. In this version, initial learning rate can be set, as in most other optimizers.

According to section 4.3 ("Effective Learning rates"), near the end of training step sizes converge to 1 which is effectively a high learning rate which would cause divergence. This occurs only near the end of the training as gradients and step sizes are small, and the epsilon constant in the numerator and denominator dominate past gradients and parameter updates which converge the learning rate to 1.

According to section 4.4("Speech Data"),where a large neural network with 4 hidden layers was trained on a corpus of US English data, ADADELTA was used with 100 network replicas.The epsilon used is 1e-6 with rho=0.95 which converged faster than ADAGRAD, by the following construction: new AdaDelta(graph, 1.0f, 0.95f, 1e-6f);