Independently Recurrent Neural Networks
Simple TensorFlow implementation of Independently Recurrent Neural Network (IndRNN): Building A Longer and Deeper RNN by Shuai Li et al.
In IndRNNs, neurons in recurrent layers are independent from each other. The basic RNN calculates the hidden state
h = act(W * input + U * state + b). IndRNNs use an element-wise vector multiplication
u * state meaning each neuron has a single recurrent weight connected to its last hidden state.
- can be used efficiently with ReLU activation functions making it easier to stack multiple recurrent layers without saturating gradients
- allows for better interpretability, as neurons in the same layer are independent from each other
- prevents vanishing and exploding gradients by regulating each neuron's recurrent weight
Copy ind_rnn_cell.py into your project.
from ind_rnn_cell import IndRNNCell # Regulate each neuron's recurrent weight as recommended in the paper recurrent_max = pow(2, 1 / TIME_STEPS) cell = MultiRNNCell([IndRNNCell(128, recurrent_max_abs=recurrent_max), IndRNNCell(128, recurrent_max_abs=recurrent_max)]) output, state = tf.nn.dynamic_rnn(cell, input_data, dtype=tf.float32) ...
Experiments in the paper
See examples/addition_rnn.py for a script reproducing the "Adding Problem" from the paper. Below are the results reproduced with the
For 100, 500 and 1000 time steps, the behavior is similar to Figure 2 in the paper. For 5000 time steps the IndRNN does not converge. My best guess is that this is due to different initialization values for the recurrent, input and softmax weights. These are not mentioned in the paper and have been found to make a big difference.
More experiments, such as Sequential MNIST, will follow in the next days.