Traditional recurrent neural networks operate as passive observers of data, unable to modify the information they learn from despite errors that may arise from suboptimal input representations. We introduce Read & Write LSTM (read-write LSTM), a new variant within the family of read & write machine learning (RW-ML) architectures that address this fundamental limitation by integrating input modification directly into the backpropagation process. Read-write LSTM establishes a dynamic feedback loop where input representations evolve alongside model weights through gradient transformation mechanisms. Our approach introduces a principled gradient scaling framework with an adaptive correction rate that carefully controls the extent of data modification, preserving data integrity while enhancing representational power. We comprehensively evaluate read-write LSTM against traditional LSTMs and state-of-the-art transformer models on the M4 competition and Numenta Anomaly Benchmark datasets, demonstrating significant improvements in forecasting accuracy. Notably, read-write LSTM consistently out-performs standard LSTM models in over 70% of time series with complex patterns and achieves superior performance on 55% of anomaly-rich datasets. Through extensive experimentation and analysis, we establish both the theoretical foundations and practical benefits of integrating data modification with neural computation, paving the way for a new generation of adaptive learning systems that actively reshape their inputs rather than merely adapting to them.

Abstract Deep learning models are widely used in multivariate time series forecasting, yet, they have high computational costs. One way to reduce this cost is by reducing data dimensionality, which involves removing unimportant or low importance information with the proper method. This work presents a study on an explainability feature selection framework composed of four methods (IMV‐LSTM Tensor, LIME‐LSTM, Average SHAP‐LSTM, and Instance SHAP‐LSTM) aimed at using the LSTM black‐box model complexity to its favour, with the end goal of improving the error metrics and reducing the computational cost on a forecast task. To test the framework, three datasets with a total of 101 multivariate time series were used, with the explainability methods outperforming the baseline methods in most of the data, be it in error metrics or computation time for the LSTM model training.

Abstract Traditional recurrent neural networks (RNNs) are essential for processing time-series data. However, they function as read-only models, lacking the ability to directly modify the data they learn from. In this study, we introduce the corrector long short-term memory (cLSTM), a Read & Write LSTM architecture that not only learns from the data but also dynamically adjusts it when necessary. The cLSTM model leverages two key components: (a) predicting LSTM’s cell states using Seasonal Autoregressive Integrated Moving Average (SARIMA) and (b) refining the training data based on discrepancies between actual and forecasted cell states. Our empirical validation demonstrates that cLSTM surpasses read-only LSTM models in forecasting accuracy across the Numenta Anomaly Benchmark (NAB) and M4 Competition datasets. Additionally, cLSTM exhibits superior performance in anomaly detection compared to hierarchical temporal memory (HTM) models.