Current Input Data Research Articles

AdaER: An adaptive experience replay approach for continual lifelong learning

Continual lifelong learning is an machine learning framework inspired by human learning, where learners are trained to continuously acquire new knowledge in a sequential manner. However, the non-stationary nature of streaming training data poses a significant challenge known as catastrophic forgetting, which refers to the rapid forgetting of previously learned knowledge when new tasks are introduced. While some approaches, such as experience replay (ER), have been proposed to mitigate this issue, their performance remains limited, particularly in the class-incremental scenario which is considered natural and highly challenging. In this paper, we present a novel algorithm, called adaptive-experience replay (AdaER), to address the challenge of continual lifelong learning. AdaER consists of two stages: memory replay and memory update. In the memory replay stage, AdaER introduces a contextually-cued memory recall (C-CMR) strategy, which selectively replays memories that are most conflicting with the current input data in terms of both data and task. Additionally, AdaER incorporates an entropy-balanced reservoir sampling (E-BRS) strategy to enhance the performance of the memory buffer by maximizing information entropy. To evaluate the effectiveness of AdaER, we conduct experiments on established supervised continual lifelong learning benchmarks, specifically focusing on class-incremental learning scenarios. The results demonstrate that AdaER outperforms existing continual lifelong learning baselines, highlighting its efficacy in mitigating catastrophic forgetting and improving learning performance.

Mini-Batch Alignment: A Deep-Learning Model for Domain Factor-Independent Feature Extraction for Wi-Fi–CSI Data

Unobtrusive sensing (device-free sensing) aims to embed sensing into our daily lives. This is achievable by re-purposing communication technologies already used in our environments. Wireless Fidelity (Wi-Fi) sensing, using Channel State Information (CSI) measurement data, seems to be a perfect fit for this purpose since Wi-Fi networks are already omnipresent. However, a big challenge in this regard is CSI data being sensitive to ‘domain factors’ such as the position and orientation of a subject performing an activity or gesture. Due to these factors, CSI signal disturbances vary, causing domain shifts. Shifts lead to the lack of inference generalization, i.e., the model does not always perform well on unseen data during testing. We present a domain factor-independent feature-extraction pipeline called ‘mini-batch alignment’. Mini-batch alignment steers a feature-extraction model’s training process such that it is unable to separate intermediate feature-probability density functions of input data batches seen previously from the current input data batch. By means of this steering technique, we hypothesize that mini-batch alignment (i) absolves the need for providing a domain label, (ii) reduces pipeline re-building and re-training likelihood when encountering latent domain factors, and (iii) absolves the need for extra model storage and training time. We test this hypothesis via a vast number of performance-evaluation experiments. The experiments involve both one- and two-domain-factor leave-out cross-validation, two open-source gesture-recognition datasets called SignFi and Widar3, two pre-processed input types called Doppler Frequency Spectrum (DFS) and Gramian Angular Difference Field (GADF), and several existing domain-shift mitigation techniques. We show that mini-batch alignment performs on a par with other domain-shift mitigation techniques in both position and orientation one-domain leave-out cross-validation using the Widar3 dataset and DFS as input type. When considering a memory-complexity-reduced version of the GADF as input type, mini-batch alignment shows hints of recuperating performance regarding a standard baseline model to the extent that no additional performance due to weight steering is lost in both one-domain-factor leave-out and two-orientation-domain-factor leave-out cross-validation scenarios. However, this is not enough evidence that the mini-batch alignment hypothesis is valid. We identified pitfalls leading up to the hypothesis invalidation: (i) lack of good-quality benchmark datasets, (ii) invalid probability distribution assumptions, and (iii) non-linear distribution scaling issues.

U-Net-LSTM: Time Series-Enhanced Lake Boundary Prediction Model

Change detection of natural lake boundaries is one of the important tasks in remote sensing image interpretation. In an ordinary fully connected network, or CNN, the signal of neurons in each layer can only be propagated to the upper layer, and the processing of samples is independent at each moment. However, for time-series data with transferability, the learned change information needs to be recorded and utilized. To solve the above problems, we propose a lake boundary change prediction model combining U-Net and LSTM. The ensemble of LSTMs helps to improve the overall accuracy and robustness of the model by capturing the spatial and temporal nuances in the data, resulting in more precise predictions. This study selected Lake Urmia as the research area and used the annual panoramic remote sensing images from 1996 to 2014 (Lat: 37°00′ N to 38°15′ N, Lon: 46°10′ E to 44°50′ E) obtained by Google Earth Professional Edition 7.3 software as the research data set. This model uses the U-Net network to extract multi-level change features and analyze the change trend of lake boundaries. The LSTM module is introduced after U-Net to optimize the predictive model using historical data storage and forgetting as well as current input data. This method enables the model to automatically fit the trend of time series data and mine the deep information of lake boundary changes. Through experimental verification, the model’s prediction accuracy for lake boundary changes after training can reach 89.43%. Comparative experiments with the existing U-Net-STN model show that the U-Net-LSTM model used in this study has higher prediction accuracy and lower mean square error.

Time series KSTAR PF superconducting coil temperature forecasting using recurrent transformer model

In this paper, we propose a KSTAR PF superconducting coil temperature forecasting model using a recurrent transformer model to protect the PF coil during operation. We developed a Recurrent Transformer model (RTransformer) that recurrently forecasts future time points using current time input data and latent variables from previous time steps based on the perceiver autoregressive model. The RTransformer model is computationally and memory efficient, requiring lower overhead than the Transformer model as it only computes forecasting points using a subset of input sequence data instead of the entire sequence of time windows. The RTransformer model was trained using the KSTAR PF coil temperature dataset obtained from the PF coil monitoring system. We compared the performance of the proposed model with Long–Short-Term memory (LSTM), Transformer model, and TransformerXL, based on R2 score, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). We also compared the inference time for each model. The experimental results showed that the proposed model outperformed other deep learning models, with lower error rates and similar inference times to the Transformer model. For most performance metrics, our proposed model scored among the top 1 and top 2 scoring models.

A Nowcasting Central Controller with Continuous RTP for Residential Device Scheduling using Swap-Based BFPSO.

Energy management plays an important role in the residential sector allowing consumers to take control over their energy consumption w.r.t. the market fluctuations. For a long time, forecasting model-based scheduling was thought as a way to mitigate the expected versus reality electricity pricing gap. However, it does not always present a working model owing to uncertainties involved around it. This paper presents a scheduling model having a Nowcasting Central Controller. This model is designed for residential devices using continuous RTP and targets on optimizing the device schedule in the current time slot as well as the subsequent time slots. It is dependent on the current input data and less on the past dataset, making it implemen at any situation. To solve the optimization problem, four variants of PSO in conjunction with swapping operation are implemented on the proposed model by considering a normalized objective function made up of two cost metrics. The results demonstrate a quickness and reduction in costs by BFPSO at each time slot. A comparison is carried out among different pricing schemes that clearly establish the effectiveness of CRTP over DAP and TOD. With CRTP performing the best of the lot, the NCC model is found to be highly adaptable and robust to sudden changes in pricing schemes.

Iterative Random Training Sampling Convolutional Neural Network for Hyperspectral Image Classification

Convolution neural network (CNN) has received considerable interest in hyperspectral image classification (HSIC) lately due to its excellent spectral-spatial feature extraction capability. To improve CNN, many approaches have been directed to exploring the infrastructure of its network by introducing different paradigms. This paper takes a rather different approach by developing an iterative CNN which extends a CNN by including a feedback system to repeatedly process the same CNN in an iterative manner. Its idea is to take advantage of a recently developed iterative training sampling spectral-spatial classification (IRTS-SSC) that allows CNN to update its spatial information of classification maps through a feedback spatial filtering system via IRTS. The resulting CNN is called iterative random training sampling CNN (IRTS-CNN) with several unique features. First, IRTS-CNN combines CNN and IRTS-SSC into one paradigm, an architecture which has never investigated in the past. Second, it implements a series of spatial filters to capture spatial information of classified data samples and further feeds this information back via an iterative process to expand the current input data cube for the next iteration. Third, it utilizes the expanded data cube to randomly re-select training samples and then to re-implement CNN iteratively. Last but not least, IRTS-CNN provides a general framework which can implement any arbitrary CNN as an initial classifier to improve its performance through an iterative process. Extensive experiments are conducted to demonstrate that IRTS-CNN indeed significantly improves CNN, specifically, when only a small size of limited training samples is used.

Modelling stochastic behaviour in simulation digital twins through neural nets

ABSTRACT In discrete event simulation (DES) models, stochastic behaviour is modelled by sampling random variates from probability distributions to determine event outcomes. However, the distribution of outcomes for an event from a real system is often dynamic and dependent on the current system state. This paper proposes the use of artificial neural networks (ANN) in DES models to determine the current distribution of each event outcome, conditional on the current model state or input data, from which random variates can then be sampled. This enables more realistic and accurate modelling of stochastic behaviour. An application is in digital twin models that aim to closely mimic a real system by learning from its past behaviour and utilising current data to predict its future. The benefits of the approach introduced in this paper are demonstrated through a realistic DES model of load-haul-dump vehicle operations in a production area of a sublevel caving mine.

Performance optimization of a reservoir computing system based on a solitary semiconductor laser under electrical-message injection.

A simple reservoir computing (RC) system based on a solitary semiconductor laser under an electrical message injection is proposed, and the performances of the RC are numerically investigated. Considering the lack of memory capacity (MC) in such a system, some auxiliary methods are introduced to enhance the MC and optimize the performances for processing complex tasks. In the pre-existing method, the input information is the current input data combined with some past input data in a weighted sum in the input layer (named as M-input). Another auxiliary method (named as M-output) is proposed to introduce the output layer for optimizing the performances of the RC system. The simulated results demonstrate that the MC of the system can be improved after adopting the auxiliary methods, and the effectiveness under adopting the M-input integrated with the M-output (named as M-both) is the most significant. Furthermore, we analyze the system performances for processing the Santa Fe time series prediction task and the nonlinear channel equalization (NCE) task after adopting the above three auxiliary methods. Results show that the M-input is the most suitable for the prediction task while the M-both is the most appropriate for the NCE task.

Spatial and temporal evaluation of storm-induced erosion modelling based on a two-dimensional field case including an artificial unvegetated research dune

The objective of this study was to evaluate the temporal and spatial performance of the XBeach model predicting storm surge-induced erosion and breaching on the basis of a two-dimensional field case. Within the scope of a physical experiment, a large-scale artificial unvegetated research dune was built at the German Baltic Sea Coastline and surveyed before, during and after a surge event to generate high-resolution digital elevation models of the temporal and spatial failure process. The measured data was used to set-up and calibrate two-dimensional XBeach models. The numerical results showed that the default settings of the latest XBeach-X release led to a very good agreement (BSS = 0.82) between the simulated and measured post-storm bed levels in the area of the research dune, while a higher agreement was observed on the overwash and inundation dominant eastern side than on the collision dominant western side. Evaluating the model accuracy during the storm surge using 44 digital elevation models of the seaside slope showed a spatially accurate prediction of the two observed breaches (±5 m) but also the occurrence of additional breaches. With respect to timing and breach width evolution, however, the results highly deviated from the observation and were found to be very sensitive to changes in grid resolution and to changes of default parameters. Exchanging the nearshore bathymetry with measurements one and two months prior to the physical experiment, where a 49% and 51% lower subaerial beach volume was observed, resulted in a significantly overestimated dune erosion (31% and 33%) and, hence, in a lower accuracy of the default model (BSS = 0.69). Overall, the XBeach model is capable of accurately predicting the evolution of storm-induced erosion during a sufficiently long and overwash/inundation-dominant surge event when current high-resolution input data is available.

Deep Neural Network for Estimation of Direction of Arrival With Antenna Array

For many object tracking systems, how to quickly and efficiently estimate the direction of arrival (DOA) of radio waves impinging on the antenna array is an urgent task. In this paper, a new efficient DOA estimation approach based on the deep neural networks (DNN) is proposed, in which a nonlinear mapping that relates the outputs of the receiving antennas with its associated DOAs is learned by using the DNN-based network. The novel network architecture is divided into two phases, the detection phase and the DOA estimation phase. Additional detection network dramatically reduces the size of the training set and the process of the training data preparation is discussed in detail. After finishing the training phase, the corresponding DOAs can be identified based on current input data during testing phase. It has been shown that the proposed method can not only achieve reasonably high DOA estimation accuracy, but also reduce the computational complexity required by traditional superresolution DOA estimation methods such as multiple signal classification (MUSIC) and estimation of signal parameters via rotation invariance (ESPRIT). The computer simulation results are performed to investigate the generalization and effectiveness of the proposed approach in different scenarios.

Sea Clutter Amplitude Prediction Using a Long Short-Term Memory Neural Network

In the marine environment, shore-based radars play an important role in military surveillance and sensing. Sea clutter is one of the main factors affecting the performance of shore-based radar. Affected by marine environmental factors and radar parameters, the fluctuation law of sea clutter amplitude is very complicated. In the process of training a sea clutter amplitude prediction model, the traditional method updates the model parameters according to the current input data and the parameters in the current model, and cannot utilize the historical information of sea clutter amplitude. It is only possible to learn the short-term variation characteristics of the sea clutter. In order to learn the long-term variation law of sea clutter, a sea clutter prediction system based on the long short-term memory neural network is proposed. Based on sea clutter data collected by IPIX radar, UHF-band radar and S-band radar, the experimental results show that the mean square error of this prediction system is smaller than the traditional prediction methods. The sea clutter suppression signal is extracted by comparing the predicted sea clutter data with the original sea clutter data. The results show that the proposed sea clutter prediction system has a good effect on sea clutter suppression.

An Integrated Approach Using Publicly Available Resources for Identifying and Characterizing Chemicals of Potential Toxicity Concern: Proof-of-Concept With Chemicals That Affect Cancer Pathways.

We developed an integrated, modular approach to predicting chemical toxicity relying on in vitro assay data, linkage of molecular targets to disease categories, and software for ranking chemical activity and examining structural features (chemotypes). We evaluate our approach in a proof-of-concept exercise to identify and prioritize chemicals of potential carcinogenicity concern. We identified 137 cancer pathway-related assays from a subset of U.S. EPA's ToxCast platforms. We mapped these assays to key characteristics of carcinogens and found they collectively assess 5 of 10 characteristics. We ranked all 1061 chemicals screened in Phases I and II of ToxCast by their activity in the selected cancer pathway-related assays using Toxicological Prioritization Index software. More chemicals used as biologically active agents (eg, pharmaceuticals) ranked in the upper 50% versus lower 50%. Twenty-three chemotypes are enriched in the top 5% (n = 54) of chemicals; these features may be important for their activity in cancer pathway-related assays. The biological coverage of the ToxCast assays related to cancer pathways is limited and short-term assays may not capture the biology of some key characteristics. Metabolism is also minimal in the assays. The ability of our approach to identify chemicals with cancer hazard is limited with the current input data, but we expect that our approach can be applied with future iterations of ToxCast and other data for improved chemical prioritization and characterization. The novel approach and proof-of-concept exercise described here for ranking chemicals for potential carcinogenicity concern is modular, adaptable, and amenable to evolving data streams.

Color segmentation in multicolor images using node‐growing self‐organizing map

AbstractThis article presents a clustering algorithm based on node‐growing self‐organizing map (NGSOM) to classify colors on color images automatically and accurately partition the regions of different colors for color measurement. Based on the CIEDE2000 criterion, pixels in a multicolor image are grouped into a number of visually distinguishable color regions in which pixel distribution information is provided as the input of the NGSOM network for further segmentation. As an unsupervised clustering algorithm, the NGSOM randomly selects two initial nodes from the input source without a predefined network structure and grows nodes according to color differences between the first best matching unit (FBMU) and the second best matching unit (SBMU) of the current input data. Unlike a traditional SOM, the NGSOM utilizes learning rates varying with the pixel distributions of major color regions. The node‐growing procedure is terminated when all the input data are examined. Compared with some commonly used color clustering algorithms, the proposed algorithm possesses a better peak signal‐to‐noise ratio (PSNR) and higher time efficiency. The NGSOM can be used for a wide range of applications, including fabric colorfastness assessment, painting conservation, and scenic identification in aerial images.

An online support vector machine for the open-ended environment

Support Vector Machine (SVM) has been widely applied in expert systems. Traditional SVM works as a closed system, which means the parameters of the SVM freeze after training is finished. This strategy leads to the SVM cannot adapt to the online learning environment and also hinders the application of the SVM in expert systems which need to be updated in an online way. Many online SVMs retrain SVM with the current support vectors and new incoming data. However, the retained system usually shows a declined performance due to losing much information about the data. To handle this problem, we propose an Online Support Vector Machine (OSVM). The OSVM includes a Representative Prototype Area (RPA), which retains the representative data of all historical data. In the RPA, each class is maintained by an Online Incremental Feature Map (OIM) which learns a suitable representative set from the stream data automatically, e.g., we do not need to predefine a compression ratio, the OIM will dynamically adapt to the current input data. Retraining is periodically conducted on the updated OIM, which can avoid a declined performance and significantly reduce the computational cost. Variety data sets from different fields are used to test our method and the experimental results demonstrate that the accuracy of OSVM is comparable to that of the state-of-the-art SVM and the speed of OSVM is much faster. Moreover, OSVM is more stable than the other methods in the open-ended learning environment, where stream data from novel classes occur. This implies the OSVM has a good application prospect in online learning expert systems.

Compact reservoir computing with a photonic integrated circuit.

Photonic reservoir computing is a new paradigm for performing high-speed prediction and classification tasks in an efficient manner. The major challenge for the miniaturization of photonic reservoir computing is the need for the use of photonic integrated circuits. Herein, we experimentally demonstrate reservoir computing using a photonic integrated circuit with a semiconductor laser and a short external cavity. We propose a method to increase the number of virtual nodes in delayed feedback using short node intervals and outputs from multiple delay times. We perform time-series prediction and nonlinear channel equalization tasks using reservoir computing with the photonic integrated circuit. We show that the photonic integrated circuit with optical feedback outperforms the photonic integrated circuit without optical feedback for prediction tasks. To enhance the memory effect we feed past input signals in the current input data and demonstrate successful performance in an n-step-ahead prediction task.

Deep Learning for an Effective Nonorthogonal Multiple Access Scheme

Nonorthogonal multiple access (NOMA) has been considered as an essential multiple access technique for enhancing system capacity and spectral efficiency in future communication scenarios. However, the existing NOMA systems have a fundamental limit: high computational complexity and a sharply changing wireless channel make exploiting the characteristics of the channel and deriving the ideal allocation methods very difficult tasks. To break this fundamental limit, in this paper, we propose a novel and effective deep learning (DL)-aided NOMA system, in which several NOMA users with random deployment are served by one base station. Since DL is advantageous in that it allows training the input signals and detecting sharply changing channel conditions, we exploit it to address wireless NOMA channels in an end-to-end manner. Specifically, it is employed in the proposed NOMA system to learn a completely unknown channel environment. A long short-term memory (LSTM) network based on DL is incorporated into a typical NOMA system, enabling the proposed scheme to detect the channel characteristics automatically. In the proposed strategy, the LSTM is first trained by simulated data under different channel conditions via offline learning, and then the corresponding output data can be obtained based on the current input data used during the online learning process. In general, we build, train and test the proposed cooperative framework to realize automatic encoding, decoding and channel detection in an additive white Gaussian noise channel. Furthermore, we regard one conventional user activity and data detection scheme as an unknown nonlinear mapping operation and use LSTM to approximate it to evaluate the data detection capacity of DL based on NOMA. Simulation results demonstrate that the proposed scheme is robust and efficient compared with conventional approaches. In addition, the accuracy of the LSTM-aided NOMA scheme is studied by introducing the well-known tenfold cross-validation procedure.

Deep Learning for Super-Resolution Channel Estimation and DOA Estimation Based Massive MIMO System

The recent concept of massive multiple-input multiple-output (MIMO) can significantly improve the capacity of the communication network, and it has been regarded as a promising technology for the next-generation wireless communications. However, the fundamental challenge of existing massive MIMO systems is that high computational complexity and complicated spatial structures bring great difficulties to exploit the characteristics of the channel and sparsity of these multi-antennas systems. To address this problem, in this paper, we focus on channel estimation and direction-of-arrival (DOA) estimation, and a novel framework that integrates the massive MIMO into deep learning is proposed. To realize end-to-end performance, a deep neural network (DNN) is employed to conduct offline learning and online learning procedures, which is effective to learn the statistics of the wireless channel and the spatial structures in the angle domain. Concretely, the DNN is first trained by simulated data in different channel conditions with the aids of the offline learning, and then corresponding output data can be obtained based on current input data during online learning process. In order to realize super-resolution channel estimation and DOA estimation, two algorithms based on the deep learning are developed, in which the DOA can be estimated in the angle domain without additional complexity directly. Furthermore, simulation results corroborate that the proposed deep learning based scheme can achieve better performance in terms of the DOA estimation and the channel estimation compared with conventional methods, and the proposed scheme is well investigated by extensive simulation in various cases for testing its robustness.

A Learning-Based Multimodel Integrated Framework for Dynamic Traffic Flow Forecasting

Accurate and timely traffic flow forecasting is essential for many intelligent transportation systems. However, it is quite challenging to develop an efficient and robust forecasting model due to the inherent randomness and large variations of traffic flow. Over the past two decades, a variety of traffic flow forecasting models have been proposed. While each model has its merits and can achieve satisfactory forecasting results under certain traffic conditions, it is difficult for a single model to deal with various conditions well. In this paper, we proposed a novel deep learning-based multimodel integration framework in order to overcome the limitations of previous methods in dealing with large variations and uncertainties of traffic flow and hence improve the forecasting accuracy. Our framework can dynamically choose an optimal model or an optimal subset of models from a set of candidate models to forecast the future traffic flow conditions according to current input data. We employ stacked autoencoder (SAE), a simple yet efficient deep learning architecture, to extract the implicit relationships hidden in the traffic flow data and employed labeled data to fine tune the parameters of the architecture. Compared with the hand-crafted features and explicable dependence relations leveraged in previous models, the features learning from SAE are more representative and hence have more powerful forecasting capability. In addition, we propose a model-driven scheme to automatically label the training data and develop three strategies to integrate multiple models. Extensive experiments performed on three typical traffic flow datasets demonstrate the proposed framework outperforms state-of-the-art models and achieves much more accurate forecasting results under large and sudden variations.

Proposing an open-source model for unconventional participation to energy planning

In this paper we present MELiSsa, a local multi-regional energy system model of a specific area: the Lombardy region. MELiSsa, implemented through an open-code modeling framework (OSeMOSYS), is built upon transparent relations and open data. Building this model is a first step towards four main goals: (i) extending the energy planning process of the region to citizens and experts usually not involved; (ii) exploiting this uncommon participation for a crowd-source development; (iii) providing a simple tool for interested local citizens to get consciousness of the technological and behavioral limits of their energy system; (iv) providing a real-case-based platform for interdisciplinary research and academic purposes possibly beyond the region boundaries. The current structure and input data of MELiSsa are presented and discussed together with a demonstrative analysis. Preliminary results show that interdisciplinary participation is enabled as an opportunity and it is needed to properly model technological dynamics as well as non-technological issues that will be relevant within the path to reach environmental, economic and social targets.

MAP: a cloud-based middleware for the provision of fine-grained compute services for mobile streaming image processing applications

Infrastructure-as-a-service offerings and the continuous evolution of wireless network technologies facilitate the integration of remote services into mobile applications. In contrast to cloud data stores, the mobile usage of cloud-based compute services is still underdeveloped. Platform-as-a-service offerings are not flexible enough to serve mobile requirements, which are performance predictability and performance adjustability based on the current connection quality. Infrastructure services are too expensive for single users, because they are normally billed at an hourly basis while the usage period is on the basis of minutes or seconds. To overcome the issues named, we propose a middleware that offers fine-grained configurable compute services for mobile users. To select a proper compute service quality according to the current needs, the remote execution performance is estimated based on the current network performance, application characteristics, and the current input data. Cloud options reflect the estimated performance for different resource types. A user may choose a cloud option for remote processing on demand. In a case study, we demonstrate the steps that are necessary to set up an application on our middleware. Furthermore, we investigate the estimation quality for a mobile image processing application.