Recent efforts to improve the interpretability of deep neural networks use saliency to characterize the importance of input features to predictions made by models. Work on interpretability using saliency-based methods on Recurrent Neural Networks (RNNs) has mostly targeted language tasks, and their applicability to time series data is less understood. In this work we analyze saliency-based methods for RNNs, both classical and gated cell architectures. We show that RNN saliency vanishes over time, biasing detection of salient features only to later time steps and are, therefore, incapable of reliably detecting important features at arbitrary time intervals. To address this vanishing saliency problem, we propose a novel RNN cell structure (input-cell attention), which can extend any RNN cell architecture.

The paper starts by showing empirically and theoretically that saliency maps generated using gradient vanishes over long sequences in LSTMs. The authors propose a modification to the LSTM cell, called LSTM with cell-attention, which can attend to all previous time steps. They show that this approach improve considerably the saliency on the input sequence. They also test their approach on the fMRI dataset of the Human Connectome Project (HCP). Originality: The proposed LSTM with cell-attention is a novel combination of well-known techniques.

Although the writing style of the paper could be improved and the presence of some perplexity about the experimental setting, the paper contributes in a valuable way to the advancement of the area. Moreover, the rebuttal helped to clarify most of the issues raised by the most negative reviewer. Overall the merits of the paper seem to overcome its drawbacks.

Recent efforts to improve the interpretability of deep neural networks use saliency to characterize the importance of input features to predictions made by models. Work on interpretability using saliency-based methods on Recurrent Neural Networks (RNNs) has mostly targeted language tasks, and their applicability to time series data is less understood. In this work we analyze saliency-based methods for RNNs, both classical and gated cell architectures. We show that RNN saliency vanishes over time, biasing detection of salient features only to later time steps and are, therefore, incapable of reliably detecting important features at arbitrary time intervals. To address this vanishing saliency problem, we propose a novel RNN cell structure (input-cell attention), which can extend any RNN cell architecture.

Recent efforts to improve the interpretability of deep neural networks use saliency to characterize the importance of input features to predictions made by models. Work on interpretability using saliency-based methods on Recurrent Neural Networks (RNNs) has mostly targeted language tasks, and their applicability to time series data is less understood. In this work we analyze saliency-based methods for RNNs, both classical and gated cell architectures. We show that RNN saliency vanishes over time, biasing detection of salient features only to later time steps and are, therefore, incapable of reliably detecting important features at arbitrary time intervals. To address this vanishing saliency problem, we propose a novel RNN cell structure (input-cell attention), which can extend any RNN cell architecture.