- TL;DR: We show how to initialize recurrent architectures with the closed-form solution of a linear autoencoder for sequences. We show the advantages of this approach compared to orthogonal RNNs.
- Abstract: Orthogonal recurrent neural networks address the vanishing gradient problem by parameterizing the recurrent connections using an orthogonal matrix. This class of models is particularly effective to solve tasks that require the memorization of long sequences. We propose an alternative solution based on explicit memorization using linear autoencoders for sequences. We show how a recently proposed recurrent architecture, the Linear Memory Network, composed of a nonlinear feedforward layer and a separate linear recurrence, can be used to solve hard memorization tasks. We propose an initialization schema that sets the weights of a recurrent architecture to approximate a linear autoencoder of the input sequences, which can be found with a closed-form solution. The initialization schema can be easily adapted to any recurrent architecture. We argue that this approach is superior to a random orthogonal initialization due to the autoencoder, which allows the memorization of long sequences even before training. The empirical analysis show that our approach achieves competitive results against alternative orthogonal models, and the LSTM, on sequential MNIST, permuted MNIST and TIMIT.
- Keywords: recurrent neural networks, autoencoders, orthogonal RNNs
- Original Pdf: pdf