Retraction Notice: Evaluating Recurrent Neural Network Architectures for Natural Language Understanding

Recurrent neural networks (RNN) are ubiquitous computing systems for sequences and multivariate time-series data. While several robust RNN architectures are known, it is unclear how to relate RNN initialization, architecture, and other hyperparameters with accuracy for a given task. In this work, we propose treating RNN as dynamical systems and correlating hyperparameters with accuracy through Lyapunov spectral analysis, a methodology designed explicitly for nonlinear dynamical systems. To address the fact that RNN features go beyond the existing Lyapunov spectral analysis, we propose to infer relevant features from the Lyapunov spectrum with an Autoencoder and an embedding of its Latent representation (AeLLE). Our studies of various RNN architectures show that AeLLE successfully correlates RNN Lyapunov spectrum with accuracy. Furthermore, the Latent representation learned by AeLLE is generalizable to novel inputs from the same task and is formed early in the process of RNN training. The latter property allows for predicting the accuracy to which RNN would converge when training is complete. We conclude that the representation of RNN through the Lyapunov spectrum, along with AeLLE, provides a novel method for the organization and interpretation of variants of RNN architectures.

Quantifying Gross Primary Production (GPP) is fundamental for understanding terrestrial carbon dynamics, particularly in forests. The overarching question we address here is whether integrating remote sensing (RS) with deep learning (DL) methodologies can enhance the estimation of daily forest GPP on a European scale.The Eddy Covariance (EC) method, although widely used to infer ecosystem-scale estimates of GPP from in situ CO2 exchange measurements, suffers from limited global coverage. When EC data are not available, RS data are often employed to estimate GPP by establishing statistical relationships with in situ observations. Recently, Machine Learning (ML) strategies, particularly involving RS and meteorological inputs, have been used for estimating GPP continuously in space and time.However, the potential of DL techniques, particularly those exploiting the sequential characteristics of time series data (i.e. recurrent neural network architectures) in estimating daily forest GPP has not been comprehensively explored. This includes their performance during photosynthetic downregulation (e.g. during climate extremes such as droughts) and an in-depth examination of the importance of the utilised features.This study presents a comparative analysis of three recurrent neural network architectures—Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs), and Long-Short Term Memory (LSTMs)—for daily forest GPP estimation across ICOS ecosystem stations. We assess their performance throughout the entire year, during the growing season, and under photosynthetic downregulation events. This assessment utilises RS inputs —optical data from Sentinel-2, Land Surface Temperature (LST) from MODIS, and radar data from Sentinel-1—, combined with potential radiation data. Furthermore, we analysed the importance of these features to provide insights into the complex interactions and dependencies when estimating GPP.Our results indicate that all three architectures yield similar accuracy for full period and growing season GPP estimations, with mean NRMSE values of 0.136 and 0.170, respectively. All models exhibit increased errors while estimating GPP during photosynthetic downregulation events and their performance varies notably, with LSTMs showing the best results (NRMSE=0.202), followed by RNNs (NRMSE=0.214), and GRUs exhibiting the least efficacy (NRMSE=0.276). Additionally, our study underscores the importance of potential radiation as a critical feature influencing the GPP response, with LST data proving particularly valuable when estimating GPP during photosynthetic downregulation events, more so than optical or radar data.These findings suggest that recurrent neural network architectures, especially LSTMs, are effective in daily GPP estimations, with slight performance decreases during climate extreme conditions. The study also reveals the efficacy of combining RS data with potential radiation for accurate forest GPP estimation, with LST data being especially crucial when estimating GPP during photosynthetic downregulation events.Our research paves the way for further exploration into recurrent models and innovative architectures, with an emphasis on overcoming the challenges in modelling climate-induced GPP extremes.

We explore the architecture of recurrent neural networks (RNNs) by studying the complexity of string sequences it is able to memorize. Symbolic sequences of different complexity are generated to simulate RNN training and study parameter configurations with a view to the network's capability of learning and inference. We compare Long Short-Term Memory (LSTM) networks and gated recurrent units (GRUs). We find that an increase in RNN depth does not necessarily result in better memorization capability when the training time is constrained. Our results also indicate that the learning rate and the number of units per layer are among the most important hyper-parameters to be tuned. Generally, GRUs outperform LSTM networks on low-complexity sequences while on high-complexity sequences LSTMs perform better.

In this paper, we propose the use of recurrent neural networks (RNNs) to develop an effective solution to two problems in electrocardiogram (ECG)-based biometrics: identification/classification and authentication. Different RNN architectures with various parameter settings were evaluated, including traditional, long short-term memory (LSTM), gated recurrent unit (GRU), unidirectional, and bidirectional networks. Unlike many existing methods, the RNN-based method does not require any feature extraction. The method was evaluated using two publicly available datasets: ECG-ID and MIT-BIH Arrhythmia (MITDB). For the identification problem, nearly 100% classification accuracy on the ECG-ID dataset was achieved, and similar results were observed for the MITDB dataset. For the authentication problem, an RNN was trained and the hidden state at the final time step was extracted to make a decision. We evaluated the effect of the training size on the equal error rate (EER), and showed that the EER drops from approximately 3.5% to nearly 0% as we increased the percentage of subjects used for training from approximately 15% to 80%.

Recurrent Neural Networks (RNNs), which are a powerful scheme for modeling temporal and sequential data need to capture long-term dependencies on datasets and represent them in hidden layers with a powerful model to capture more information from inputs. For modeling long-term dependencies in a dataset, the gating mechanism concept can help RNNs remember and forget previous information. Representing the hidden layers of an RNN with more expressive operations (i.e., tensor products) helps it learn a more complex relationship between the current input and the previous hidden layer information. These ideas can generally improve RNN performances. In this paper, we proposed a novel RNN architecture that combine the concepts of gating mechanism and the tensor product into a single model. By combining these two concepts into a single RNN, our proposed models learn long-term dependencies by modeling with gating units and obtain more expressive and direct interaction between input and hidden layers using a tensor product on 3-dimensional array (tensor) weight parameters. We use Long Short Term Memory (LSTM) RNN and Gated Recurrent Unit (GRU) RNN and combine them with a tensor product inside their formulations. Our proposed RNNs, which are called a Long-Short Term Memory Recurrent Neural Tensor Network (LSTMRNTN) and Gated Recurrent Unit Recurrent Neural Tensor Network (GRURNTN), are made by combining the LSTM and GRU RNN models with the tensor product. We conducted experiments with our proposed models on word-level and character-level language modeling tasks and revealed that our proposed models significantly improved their performance compared to our baseline models.

In this article, we propose a new behavioral thermal modeling method for fast building performance analysis, which is critical for energy-efficient smart building control and management. The new approach is based on two recurrent neutral network architecture to obtain the compact nonlinear thermal models for complicated building. We start with a more realistic building simulation program, EnergyPlus, from Department of Energy, to model some practical buildings such as office buildings and data centers. EnergyPlus can model the various time-series inputs to a building such as ambient temperature, heating, ventilation, and air-conditioning (HVAC) inputs, power consumption of electronic equipment, lighting and number of occupants in a room sampled in each hour and produce resulting temperature traces of zones (rooms). In this work, we apply two recurrent neural network (RNN) architectures to build the non-linear compact thermal model of the building: one is non-linear state-space RNN architecture (NLSS), which has global feedbacks, and the other one is Elman's RNN architecture (ELNN), which has local feedbacks in each layer. We give a simple formula to calculate the RNN layer number, layer size to configure RNN architecture to avoid overfitting and underfitting problems. A cross-validation based training technique is further applied to improve predictable accuracy of models. Experimental results from a case study of three buildings show that ELNN and NLSS can both build very accurate building thermal models for the 2-zone and 5-zone building cases: both of them have average errors from around 1% to 1.5% for the two buildings. For the more complex 6-zone building case, ELNN outperforms NLSS with maximum errors 16% against 23%. But both methods have 2.2% average errors.

MOOCs (Massive Open Online Courses) are definitely one of the best approach to support the international agenda about inclusive and equitable education and lifelong learning opportunities for all (SDG4) [1]. A great deal universities and institutions offer valuable free courses to their numerous students and to people around the word through MOOC platforms. However, because of huge number of learners and data generated, learner’s behaviour in those platforms remain a kind of black box for learners themselves and for courses instructors who are supposed to tutor or monitor learners in the learning process. Therefore, learner do not receive sufficient support from instructors and from their peers, during the learning process [2]. This is one the main reasons that lead to high dropout, low completion and success rate observed in the MOOCs. Many research work have suggested descriptive, predictive and prescriptive models to address this issue, but most of these models focus on predicting dropout, completion and/or success, and do not generally pay enough attention to one of the key step (learner behaviour), that comes before, and can explain dropping out and failure. Our research aims to develop a deep learning model to predict learner behaviour (learner interactions) in the learning process, in order to equip learners and course instructors with insight understanding of the learner behaviour in the learning process. This specific paper will focus on analysing relevant algorithms to develop such model. For this analysis, we used data from UNESCO-IICBA (International Institute for Capacity Building in Africa) MOOC platform, designed for teacher training in Africa, and then we examine many types of features: geographical, social behavioural and learning behavioural features. Learner’s behaviour being a time series Big data, we built the predictive model using Deep Learning algorithms for better performance and accuracy (Thanks to the power of deep learning) compared to baseline Machine learning algorithms. Time series data is best handled by recurrent neural networks (RNN) [3], so, we choose RNN and implemented/tested the three main architectures of RNN: Simple RNNs, GRU (Gated Recurrent Unit) RNNs and LSTM (Long short-term memory) RNNs. The models were trained using L2 Regularization, based on the predictions results, we unexpectedly found model with simple RNNs produced the best performance and accuracy on the dataset used than the other RNN architectures. We had couple of observations, example: we saw a correlation between video viewing and quiz behaviour and the participation of the learner to the learning process. This observation could allow teachers to provide meaningful support and guidance to at risk or less active students. We also observed that, the shorter the video or the quiz, the more the viewer. We conclude that we could use learner video or quiz viewing behaviour to predict his behaviour concerning other MOOC contents. These results suggest the need of deeper research on educational video and educational quiz designing for MOOCs.

Forecasting longevity is essential in multiple research and policy areas, including the pricing of life insurance contracts, the valuation of capital market solutions for longevity risk management, and pension policy. This paper empirically investigates the predictive performance of Recurrent Neural Networks (RNN) with Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) architectures in jointly modeling and multivariate time series forecasting of age-specific mortality rates at all ages. We fine-tune the three hidden layers GRU and LSTM model’s hyperparameters for time series forecasting and compare the model’s forecasting accuracy with that produced by traditional Generalised Age-Period-Cohort (GAPC) stochastic mortality models. The empirical results suggest that the two RNN architectures generally outperform the GAPC models investigated in both the male and female populations, but the results are sensitive to the accuracy criteria. The empirical results also show that the RNN-GRU network slightly outperforms the RNN with an LSTM architecture and can produce mortality schedules that capture relatively well the dynamics of mortality rates across age and time. Further investigations considering other RNN architectures, calibration procedures, and sample datasets are necessary to confirm the superiority of RNN in forecasting longevity.

Applications like smart homes, security surveillance, healthcare, and workplace safety now depend heavily on Human Activity Recognition (HAR) and monitoring. Automated systems can respond quickly, stop dangerous situations, and provide individualized services when they can precisely identify and categorize human activity. The intricate temporal dependencies present in human motion are frequently difficult to capture by conventional HAR techniques that rely on hand-crafted features and shallow learning models [1]. This paper investigates the effectiveness of a pure Recurrent Neural Network (RNN) architecture for activity recognition using smartphone sensor data, despite the fact that deep learning models like CNNs and LSTMs have dominated recent HAR research. Without the need for manually created features, we suggest a simplified RNN-based framework that can identify temporal dependencies in unprocessed accelerometer and gyroscope data. Pure RNNs can attain competitive accuracy, as shown by experiments on the UCI HAR and WISDM datasets, which yielded results of 92.4% and 90.7%, respectively. The results imply that, when properly optimized, RNNs despite their simplicity remain feasible for HAR tasks.

Purpose: This study aims to predict the closing price of the EUR/JPY currency pair in the forex market using recurrent neural network (RNN) architectures, namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), with the incorporation of Bidirectional layers. Methods: The dataset comprises hourly price data obtained from Yahoo Finance and pre-processed accordingly. The data is divided into training and testing sets, and time series sequences are constructed for input into the models. The RNN, LSTM, and GRU models are trained using the Adam optimization algorithm with the mean squared error (MSE) loss metric. Results: Results indicate that the LSTM model, particularly when coupled with Bidirectional layers, exhibits superior predictive performance compared to the other models, as evidenced by lower MSE values. Conclusions: Therefore, the LSTM model with Bidirectional layers is the most effective in predicting the EUR/JPY currency pair's closing price in the forex market. These findings offer valuable insights for practitioners and researchers involved in financial market prediction and neural network modelling

Emotion classification using EEG (Electroencephalogram) brain signals has recently garnered significant interest due to its potential applications in fields such as mental health, human-computer interaction, and affective computing. The three prominent recurrent neural network (RNN) architectures GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory), and BiLSTM (Bidirectional Long Short-Term Memory) are thoroughly compared in this study with a focus on classifying emotions. The study examines how well these RNN variations capture temporal dependencies and patterns within EEG data to precisely discriminate between various emotional states. Based on classification accuracy and training effectiveness, these model's performance will be assessed. The results will reveal the most suitable architecture in terms of accuracy and robustness for EEG-based emotion classification. The findings provide valuable insights into choosing the appropriate RNN architecture for emotion classification tasks, advancing the understanding of AI's potential in interpreting human emotions through EEG signals.

Recently, emotion recognition began to be implemented in the industry and human resource field. In the time we can perceive the emotional state of the employee, the employer could gain benefits from it as they could improve the quality of decision makings regarding their employee. Hence, this subject would become an embryo for emotion recognition tasks in the human resource field. In a fact, emotion recognition has become an important topic of research, especially one based on physiological signals, such as EEG. One of the reasons is due to the availability of EEG datasets that can be widely used by researchers. Moreover, the development of many machine learning methods has been significantly contributed to this research topic over time. Here, we investigated the classification method for emotion and propose two models to address this task, which are a hybrid of two deep learning architectures: One-Dimensional Convolutional Neural Network (CNN-1D) and Recurrent Neural Network (RNN). We implement Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) in the RNN architecture, that specifically designed to address the vanishing gradient problem which usually becomes an issue in the time-series dataset. We use this model to classify four emotional regions from the valence-arousal plane: High Valence High Arousal (HVHA), High Valence Low Arousal (HVLA), Low Valence High Arousal (LVHA), and Low Valence Low Arousal (LVLA). This experiment was implemented on the well-known DEAP dataset. Experimental results show that proposed methods achieve a training accuracy of 96.3% and 97.8% in the 1DCNN-GRU model and 1DCNN-LSTM model, respectively. Therefore, both models are quite robust to perform this emotion classification task.

In this report, a deep learning network architecture, which combined with the architecture of convolutional neural network, the architecture of recurrent neural network is proposed and implemented to do the text classification with the pre-trained word vectors. The proposed method set the word vector be as a static lookup table without updating, and the network still can ignore the noise which caused by missing words. The experimental results show that the accuracy of this study is consistent with the accuracy of other studies. It is shown the feasibility of this architecture. And has the following advantages: the accuracy rate of this architecture is higher than that of recurrent neural network only; compared with the convolutional neural network, the accuracy results are more stable; and less epoch is used to get stable results. But the shortcoming of this proposed architecture is that the training time will consumed much time.

To reduce the computational complexity and improve the performance of the recurrent wavelet neural network (RWNN), a novel modular recurrent neural network based on the pipelined architecture (PRWNN) with low computational complexity is presented in this paper. Its modified adaptive real-time recurrent learning (RTRL) algorithm is derived on the gradient descent approach. The PRWNN comprises a number of RWNN modules that are cascaded in a chained form and inherits the modular architectures of the pipelined recurrent neural network (PRNN) proposed by Haykin and Li. Since those modules of the PRWNN can be performed simultaneously in a pipelined parallelism fashion, it would result in a significant improvement in computational efficiency. And the performance of the PRWNN can be also further improved. Computer simulations have demonstrated that the PRWNN provides considerably better performance compared to the single RWNN model for nonlinear dynamic system identification.

Funding Information: V. CONCLUSIONS This work proposed four classification models based on LSTM RNNs to classify SIP dialogs. The detection probability was evaluated based on experimental data. To detect abnormal SIP dialogs, we have adopted classification features computed from the output of the LSTM RNN model and two different classification schemes were proposed. A semi-supervised scheme is shown to reach higher performance, achieving a detection probability of 99.45%, thus confirming the effective utility of the proposed methodology to detect abnormal SIP sequences in a short period of time. ACKNOWLEDGEMENTS This work was funded by Fundac¸ão para a Ciência e Tecnologia, under the projects InfoCent-IoT (PTDC/EEI-TEL/30433/2017), CoSHARE (PTDC/EEI-TEL/30709/2017), and RFSense (UIDB/50008/2020).