Recurrent neural networks can learn complex transduction problems that require maintaining and actively exploiting a memory of their inputs. Such models traditionally consider memory and input-output functionalities indissolubly entangled. We introduce a novel recurrent architecture based on the conceptual separation between the functional input-output transformation and the memory mechanism, showing how they can be implemented through different neural components. By building on such conceptualization, we introduce the Linear Memory Network, a recurrent model comprising a feedforward neural network, realizing the non-linear functional transformation, and a linear autoencoder for sequences, implementing the memory component. The resulting architecture can be efficiently trained by building on closed-form solutions to linear optimization problems. Further, by exploiting equivalence results between feedforward and recurrent neural networks we devise a pretraining schema for the proposed architecture. Experiments on polyphonic music datasets show competitive results against gated recurrent networks and other state of the art models.
The Long Short-Term Memory (LSTM) layer is an important advancement in the field of neural networks and machine learning, allowing for effective training and impressive inference performance. LSTM-based neural networks have been successfully employed in various applications such as speech processing and language translation. The LSTM layer can be simplified by removing certain components, potentially speeding up training and runtime with limited change in performance. In particular, the recently introduced variants, called SLIM LSTMs, have shown success in initial experiments to support this view. Here, we perform computational analysis of the validation accuracy of a convolutional plus recurrent neural network architecture using comparatively the standard LSTM and three SLIM LSTM layers. We have found that some realizations of the SLIM LSTM layers can potentially perform as well as the standard LSTM layer for our considered architecture.
Modelling long-term dependencies is a challenge for recurrent neural networks. This is primarily due to the fact that gradients vanish during training, as the sequence length increases. Gradients can be attenuated by transition operators and are attenuated or dropped by activation functions. Canonical architectures like LSTM alleviate this issue by skipping information through a memory mechanism. We propose a new recurrent architecture (Non-saturating Recurrent Unit; NRU) that relies on a memory mechanism but forgoes both saturating activation functions and saturating gates, in order to further alleviate vanishing gradients. In a series of synthetic and real world tasks, we demonstrate that the proposed model is the only model that performs among the top 2 models across all tasks with and without long-term dependencies, when compared against a range of other architectures.
Vanishing long-term gradients are a major issue in training standard recurrent neural networks (RNNs), which can be alleviated by long short-term memory (LSTM) models with memory cells. However, the extra parameters associated with the memory cells mean an LSTM layer has four times as many parameters as an RNN with the same hidden vector size. This paper addresses the vanishing gradient problem using a high order RNN (HORNN) which has additional connections from multiple previous time steps. Speech recognition experiments using British English multi-genre broadcast (MGB3) data showed that the proposed HORNN architectures for rectified linear unit and sigmoid activation functions reduced word error rates (WER) by 4.2% and 6.3% over the corresponding RNNs, and gave similar WERs to a (projected) LSTM while using only 20%--50% of the recurrent layer parameters and computation.