Gated Recurrent Unit (GRU)

                               Gated Recurrent Unit (GRU)

Outline          

                  Background

                  GRU Network

                  GRU vs. LSTM

                  Experiment

                  References

Background

A gated recurrent unit (GRU) was proposed by Cho et al. [2014] to make each recurrent unit to adaptively capture dependencies of different time scales.

Solving problems existed in RNN: Gradient Vanishing.

Example:

GRU Network

GRU vs. LSTM

Code Example:

import tensorflow as tf

x = tf.constant([[1]], dtype = tf.float32)

…

state0_lstm = lstm_cell.zero_state(1,dtype=tf.float32)

output,state = lstm_cell(x,state0_lstm)

state0_gru = gru_cell.zero_state(1,dtype=tf.float32)

output2,state2 = gru_cell(x,state0_gru)

with tf.Session() as sess:

sess.run(init)

print(sess.run(output))

print(sess.run(state))

print(sess.run(output2))

print(sess.run(state2))

Experiment

References

1. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

2. Learned-norm pooling for deep feedforward and recurrent neural networks

3. Long short-term memory