Shorter Training Time Research Articles

Co-phase errors sensing method for Golay3 telescope system via a transfer network.

Optical sparse-aperture systems face severe challenges, including detecting and correcting co-phase errors. In this study, a search framework based on fine tuning a pre-trained network is proposed to analyze the co-phase errors of a Golay3 telescope system. Based on this, an error compensation control system is established. First, a hash-like binary code is created by fine-tuning the pre-trained model. Secondly, a pre-trained network is used to extract the deep features of the image, and an index database is built between the image features and the corresponding co-phase error values. Finally, the Top 1-ranked features and corresponding co-phase error values are returned through the hash-like binary code hierarchical deep search database to provide driving data for the error correction system. Numerical simulations and experimental results verify the method's validity. The experimental results show that the correction system works well when the dynamic piston is [-5,5]λ, and the tilt error range is [-15,15]µr a d. Compared with existing detection methods, this method does not require additional optical components, has a high correction accuracy, and requires a short training time. Furthermore, it can be used to detect piston and tilt errors simultaneously.

Gray whale detection in satellite imagery using deep learning

AbstractThe combination of very high resolution (VHR) satellite remote sensing imagery and deep learning via convolutional neural networks provides opportunities to improve global whale population surveys through increasing efficiency and spatial coverage. Many whale species are recovering from commercial whaling and face multiple anthropogenic threats. Regular, accurate population surveys are therefore of high importance for conservation efforts. In this study, a state‐of‐the‐art object detection model (YOLOv5) was trained to detect gray whales (Eschrichtius robustus) in VHR satellite images, using training data derived from satellite images spanning different sea states in a key breeding habitat, as well as aerial imagery collected by unoccupied aircraft systems. Varying combinations of aerial and satellite imagery were incorporated into the training set. Mean average precision, whale precision, and recall ranged from 0.823 to 0.922, 0.800 to 0.939, and 0.843 to 0.889, respectively, across eight experiments. The results imply that including aerial imagery in the training data did not substantially impact model performance, and therefore, expansion of representative satellite datasets should be prioritized. The accuracy of the results on real‐world data, along with short training times, indicates the potential of using this method to automate whale detection for population surveys.

An Actinic Keratosis Auxiliary Diagnosis Method Based on an Enhanced MobileNet Model.

Actinic keratosis (AK) is a common precancerous skin lesion with significant harm, and it is often confused with non-actinic keratoses (NAK). At present, the diagnosis of AK mainly depends on clinical experience and histopathology. Due to the high difficulty of diagnosis and easy confusion with other diseases, this article aims to develop a convolutional neural network that can efficiently, accurately, and automatically diagnose AK. This article improves the MobileNet model and uses the AK and NAK images in the HAM10000 dataset for training and testing after data preprocessing, and we performed external independent testing using a separate dataset to validate our preprocessing approach and to demonstrate the performance and generalization capability of our model. It further compares common deep learning models in the field of skin diseases (including the original MobileNet, ResNet, GoogleNet, EfficientNet, and Xception). The results show that the improved MobileNet has achieved 0.9265 in accuracy and 0.97 in Area Under the ROC Curve (AUC), which is the best among the comparison models. At the same time, it has the shortest training time, and the total time of five-fold cross-validation on local devices only takes 821.7 s. Local experiments show that the method proposed in this article has high accuracy and stability in diagnosing AK. Our method will help doctors diagnose AK more efficiently and accurately, allowing patients to receive timely diagnosis and treatment.

Partial Discharge Pattern Recognition Method Based on Transfer Learning and DenseNet Model

With the development of intelligent sensing technology, a large amount of partial discharge (PD) time-domain waveform images are generated in the on-site detection of gas insulated switchgear (GIS) partial discharge. Traditional pattern recognition methods are mostly aimed at structured data and cannot directly identify defect types of such data.At the same time, the deep learning method for GIS PD pattern recognition is generally faced with the problem of small samples.In order to solve the above problems, this paper proposes a PD pattern recognition method based on transfer learning and DenseNet model.Firstly, the time-domain waveform images are processed by image enhancement, normalization, image compression and other image processing techniques.The FDTD method was used to simulate GIS PD, and the time domain waveform image database of four PD defects are established.Using CNN and transfer learning, the recognition accuracy of the model is increased to 95%, with better robustness.The recognition performance of different CNN structures is studied. The results show that DenseNet model has higher accuracy than other structures and shorter training time. This study can be used to diagnose the insulation status of GIS equipment in-site.

Feasibility and effect of 3/7 strength training method in cardiac rehabilitation: preliminary results

Abstract Funding Acknowledgements Type of funding sources: Foundation. Main funding source(s): Erasmus Hospital. Background Cardiovascular rehabilitation (CR) is a type 1A intervention for cardiovascular disease. Among the various exercise modalities, resistance training (RT) is recommended for improving cardiorespiratory fitness (volume of oxygen consumption [VO2] peak). However, different modalities of RT coexist and which modalities should be preferred is still a matter of discussion. A RT method, called 3/7, allowed in healthy young subject greater strength gains with a shorter training time. However, the feasibility, the safety and its effects have never been studied in CR. Purpose The aim was to assess the feasibility, safety and efficiency of the 3/7 method in CR to improve VO2 peak. Methods In this single center randomized clinical trial, participants underwent 12 weeks of supervised training. Intervention: Group 3/7 consisting of 5 incremental sets of 3 to 7 repetitions with 15 s of inter-set intervals of RT on leg press; leg extension, leg curl, triceps press machines. Group 3X9 (usual care) of 3 sets of 9 repetitions with 1 min of inter set intervals on the same machine and at the same intensity (70% of 1RM). Patients completed 3 sessions per week; endurance training is the same in both groups. The primary outcome was a change in VO2 peak during the cardiopulmonary exercise test from baseline to 12 weeks. Secondary outcomes were isokinetic leg strength at 60°/sec and feasibility. Results Between 2019 and 2022 101 patients (92 male) were enrolled and randomized to 3/7 (n=49) or 3x9 RT (n=52). The mean age was 62y±8 (range 34 – 85). A total of 65 participants completed testing at 12 weeks for the primary outcome, including 32 in the 3/7 group and 33 in the 3X9 group. There were similar improvements from baseline between groups, VO2 peak by 19% in 3/7 group with 17% in the 3X9 group (mean [SD] oxygen uptake: 3/7, 3.6 [4.1] mL/kg/min; 3X9, 3.3 [4.5] mL/kg/min; P = .4). 3/7 improved quadriceps strength by 12% compared with 11% in the 3X9 group (mean [SD] isokinetic leg strength: 3/7, 23 [21] N/m; 3X9, 10 [22] N/m; P = .01). We had more dropouts in the usual care group compared to the 3/7 group (19 vs 17). 3/7 group had high feasibility scores and low rates of withdrawal due to serious adverse events (1). Most dropouts were due to other conditions (Lost of follow-up [n=11], COVID-19 [n=6], other medical problems [n=6], returns to work [n=5], withdrew [n=4]). Conclusion(s) In this randomized clinical trial, a 12-week RT of the 3/7 method coupled with endurance training improved VO2 peak to the same level than the usual care in cardiac rehabilitation, but more leg strength was observed. The 3/7 method may reduce the number of dropouts compared to usual RT program because the time training is reduced for the same volume of work. Implications: These findings support inclusion of 3/7 method in cardiac rehabilitation programs as an adjunct or alternative modality to usual care.

Exciton diffusion in amorphous organic semiconductors: Reducing simulation overheads with machine learning.

Simulations of exciton and charge hopping in amorphous organic materials involve numerous physical parameters. Each of these parameters must be computed from costly abinitio calculations before the simulation can commence, resulting in a significant computational overhead for studying exciton diffusion, especially in large and complex material datasets. While the idea of using machine learning to quickly predict these parameters has been explored previously, typical machine learning models require long training times, which ultimately contribute to simulation overheads. In this paper, we present a new machine learning architecture for building predictive models for intermolecular exciton coupling parameters. Our architecture is designed in such a way that the total training time is reduced compared to ordinary Gaussian process regression or kernel ridge regression models. Based on this architecture, we build a predictive model and use it to estimate the coupling parameters which enter into an exciton hopping simulation in amorphous pentacene. We show that this hopping simulation is able to achieve excellent predictions for exciton diffusion tensor elements and other properties as compared to a simulation using coupling parameters computed entirely from density functional theory. This result, along with the short training times afforded by our architecture, shows how machine learning can be used to reduce the high computational overheads associated with exciton and charge diffusion simulations in amorphous organic materials.

V-A3tS: A rapid text steganalysis method based on position information and variable parameter multi-head self-attention controlled by length

Text length varies in social networks, such as IMDB long text, Twitter short text, and long–short mixed text. For these complex situations, the time series, convolution, or fine-tuned BERT models are used in existing text steganalysis methods almost. However, these methods do not simultaneously consider higher detection accuracy and lower training time at the same time. To alleviate this dilemma, this paper proposes a novel text steganalysis method. First, the proposed method maps words into a semantic space containing position information. Second, a variable parameter attention layer scaled appropriately according to text length is designed, it achieves the purpose that the entire parameter amount of the model is not redundant and can ensure effective detection. Finally, the steganalysis features are enhanced by the residual linear layer. For long, short, and mixed text datasets, comparing experiments show that the proposed method has higher detection accuracy, fewer parameters, and shorter training time than existing methods. Among them, the advantage of this method is more obvious for long and mixed texts.

Pathologists aren't pigeons: exploring the neural basis of visual recognition and perceptual expertise in pathology.

Visual (perceptual) reasoning is a critical skill in many medical specialties, including pathology, diagnostic imaging, and dermatology. However, in an ever-compressed medical curriculum, learning and practicing this skill can be challenging. Previous studies (including work with pigeons) have suggested that using reward-feedback-based activities, novices can gain expert levels of visual diagnostic accuracy in shortened training times. But is this level of diagnostic accuracy a result of image recognition (categorization) or is it the acquisition of diagnostic expertise? To answer this, the authors measured electroencephalographic data (EEG) and two components of the human event-related brain potential (reward positivity and N170) to explore the nature of visual expertise in a novice-expert study in pathology visual diagnosis. It was found that the amplitude of the reward positivity decreased with learning in novices (suggesting a decrease in reliance on feedback, as in other studies). However, this signal remained significantly different from the experts whose reward positivity signal did not change over the course of the experiment. There were no changes in the amplitude of the N170 (a reported neural marker of visual expertise) in novices over time. Novice N170 signals remained statistically and significantly lower in amplitude compared to experts throughout task performance. These data suggest that, while novices gained the ability to recognize (categorize) pathologies through reinforcement learning as quantified by the change in reward positivity, increased accuracy, and decreased time for responses, there was little change in the neural marker associated with visual expertise (N170). This is consistent with the multi-dimensional and complex nature of visual expertise and provides insight into future training programs for novices to bridge the expertise gap.

Prediction of PM10 Concentration in Malaysia Using K-Means Clustering and LSTM Hybrid Model

Following the rapid development of various industrial sectors, air pollution frequently occurs in every corner of the world. As a dominant pollutant in Malaysia, particulate matter PM10 can cause highly detrimental effects on human health. This study aims to predict the daily average concentration of PM10 based on the data collected from 60 air quality monitoring stations in Malaysia. Building a forecasting model for each station is time-consuming and unrealistic; therefore, a hybrid model that combines the k-means clustering technique and the long short-term memory (LSTM) model is proposed to reduce the number of models and the overall model training time. Based on the training set, the stations were clustered using the k-means algorithm and an LSTM model was built for each cluster. Then, the prediction performance of the hybrid model was compared with the univariate LSTM model built independently for each station. The results show that the hybrid model has a comparable prediction performance to the univariate LSTM model, as it gives the relative percentage difference (RPD) less than or equal to 50% based on at least two accuracy metrics for 43 stations. The hybrid model can also fit the actual data trend well with a much shorter training time. Hence, the hybrid model is more competitive and suitable for real applications to forecast air quality.

A PCA-IGRU Model for Stock Price Prediction

<p>Accurate stock price prediction is significant for investors to avoid risks and improve the return on investment. Stock price prediction is a typical nonlinear time-series problem, which many factors affect. Still, too much analysis of influencing factors will lead to input redundancy and a large amount of computation in the model. Although the stock prediction model based on Recurrent Neural Network (RNN) has a good prediction effect, it has the problem of oversaturation. This paper proposes a prediction model of stock closing price based on Principal Component Analysis (PCA) and Improved Gated Recurrent Unit (IGRU), PCA-IGRU. PCA can reduce the redundancy of input information without destroying the correlation of original data, thus reducing the time of model training and prediction. IGRU is an improved Gated Recurrent Unit (GRU) model, which prevents oversaturation by introducing the Anti-oversaturation Conversion Module (ACM) and enhances the sensitivity of model learning. This paper selects the stock trading data of the Shanghai Composite Index (SCI) of China as experimental data. The PCA-IGRU is compared with seven baseline models. The experimental results show that the model has better prediction accuracy and shorter training time.</p> <p> </p>

Convolutional neural network optimized by differential evolution for electrocardiogram classification.

The Coronavirus disease 2019, or COVID-19, has shifted the medical paradigm from face-to-face to telehealth. Telehealth has become a vital resource to contain the virus spread and ensure the continued care of patients. In terms of preventing cardiovascular diseases, automating electrocardiogram (ECG) classification is a promising telehealth intervention. The healthcare service ensures that patient care is appropriate, comfortable, and accessible. Convolutional neural networks (CNNs) have demonstrated promising results in ECG categorization, which require high accuracy and short training time to ensure healthcare quality. This study proposes a one-dimensional-CNN (1D-CNN) arrhythmia classification based on the differential evolution (DE) algorithm to optimize the accuracy of ECG classification and training time. The performance of 1D-CNNs of different activation functions are optimized based on the standard DE algorithm. Finally, based on MIT-BIH and SCDH arrhythmia databases, the performances of optimized and unoptimized 1D-CNN are compared and analysed. Results show that the 1D-CNN optimized by the DE has higher accuracy in heartbeats classification. The optimized 1D-CNN improves from 97.6% to 99.5% on MIT-BIH and from 80.2% to 88.5% on SCDH. Therefore, the optimized 1D-CNN shows improvements of 1.9% and 8.3% in the two datasets, respectively. In addition, compared with the unoptimized 1D-CNN based on the same parameter settings, the optimized 1D-CNN has less training time. Under the conditions of ReLU function and 10 epochs, the training takes 9.22s on MIT-BIH and 10.35s on SCDH, reducing training time by 67.2% and 64.2%, respectively.

A Study of Forest Phenology Prediction Based on GRU Models

Investigating forest phenology prediction is a key parameter for assessing the relationship between climate and environmental changes. Traditional machine learning models are not good at capturing long-term dependencies due to the problem of vanishing gradients. In contrast, the Gated Recurrent Unit (GRU) can effectively address the problem of vanishing gradients and allow the neural network to capture longer-range dependencies. In this study, an optical camera was used as experimental equipment to obtain forest images. The absolute greenness index (GEI) data of the region of interest (ROI) in the images were calculated to fit the seasonal variation curve of forest phenology. The GRU neural network model was introduced to train and analyze the GEI data, and the performance of the GRU neural network was evaluated using the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) methods. Finally, the model was used to predict the trend of GEI data for the next 60 days. The results showed that: (1) In terms of training and predicting forest phenology, the GRU model was validated using histograms and autocorrelation graphs, which indicated that the distribution of predicted data was consistent with the trend of actual data, the GRU model data was feasible, and the model was stable. (2) The MSE values of the GRU model at rain-fed-CK (preset point 1), sufficient drip irrigation-DIFI (preset point 3), and sufficient furrow irrigation-BIFI (preset point 5) were 9.055 × 10−5, 12.91 × 10−5, and 8.241 × 10−5, respectively. The RMSE values were 9.516 × 10−3, 11.36 × 10−3, and 7.313 × 10−3, respectively. The MAE values were 7.174 × 10−3, 8.241 × 10−3, and 5.351 × 10−3, respectively. These results indicate that the overall performance of the GRU model was good. (3) The predicted GEI data for the next 60 days showed a trend consistent with actual changes in GEI data, as demonstrated by the GRU model. The GRU model has become the preferred method for phenological prediction due to its simple internal structure and relatively short training time. Results show that the GRU model can achieve forest phenological change prediction and can reveal in-depth insights into future forest growth and climate change, providing a theoretical basis for the application of forest phenological prediction.

Learning position information from attention: End-to-end weakly supervised crack segmentation with GANs

Despite the impressive progress of fully supervised crack segmentation, the tedious pixel-level annotation restricts its general application. Weakly supervised crack segmentation with image-level labels has received increasing attention due to the easily accessible annotation. However, the current methods are mainly based on class activation mapping (CAM) and fail to obtain the accurate crack position information directly, resulting in the complex training steps and poor segmentation performance. For the efficient tasks of weakly supervised crack segmentation, this paper proposes a novel end-to-end weakly supervised crack segmentation method termed RepairerGAN, which can directly obtain the crack segmentation result with the category information only. The proposed RepairerGAN decouples the image-to-image translation model of two different image domains into a semantic translation module and a position extraction module and uses the attention mechanism to extract the crack position information as the segmentation result. In the simple weakly supervised segmentation task based on METU crack dataset, the performance of RepairerGAN only needs a training time equal to 13.3% of that of the best performing ScoreCAM. In the complex task based on Combined crack dataset, the performance of RepairerGAN (F1 of 72.63% and IoU of 61.37%) with shorter training time is significantly ahead of the best performing ScoreCAM (F1 of 44.43% and IoU of 33.32%).

Improving the Performance of Deep Q-learning in Games Pong and Ms. Pacman

Unlike TD-gammon architecture, deep Q-learning algorithm combines deep learning and reinforcement learning, and it has achieved outstanding results in many Atari games. This study primarily focuses on two games, Pong and Ms. Pacman. It explores several approaches to improve the performance of deep Q-network (DQN). Based on the data obtained, while DQN displays a high-level performance in the simple Atari game Pong, it struggles a bit when learning the more complex game Ms. Pacman, leading to diverged loss. This under-performance may be partly explained by the shorter training time than the original paper due to limited computational resource. Given sufficient time of training and exploring, the model is believed to eventually converge once it identifies the optimal combination of hyperparameters.

Artificial intelligence-based prediction of indoor bioaerosol concentrations from indoor air quality sensor data

Exposure to bioaerosols in indoor environments, especially public venues that have a high occupancy and poor ventilation, is a serious public health concern. However, it remains challenging to monitor and determine real-time or predict near-future concentrations of airborne biological matter. In this study, we developed artificial intelligence (AI) models using physical and chemical data from indoor air quality sensors and physical data from ultraviolet light-induced fluorescence observations of bioaerosols. This enabled us to effectively estimate the bioaerosol (bacteria-, fungi- and pollen-like particle) and 2.5-µm and 10-µm particulate matter (PM2.5 and PM10) on a real-time and near-future (≤60 min) basis. Seven AI models were developed and evaluated using measured data from an occupied commercial office and a shopping mall. A long short-term memory model required a relatively short training time and gave the highest prediction accuracy of ∼ 60 %–80 % for bioaerosols and ∼ 90 % for PM on the testing and time series datasets from the two venues. This work demonstrates how AI-based methods can leverage bioaerosol monitoring into predictive scenarios that building operators can use for improving indoor environmental quality in near real-time.

An error based adaptive learning rate stochastic gradient descent algorithm in convolutional neural network

If the learning rate of convolutional neural network (CNN) is set improperly, the efficiency and accuracy of the algorithm will be greatly affected. To solve this problem, a learning rate adaptive algorithm is proposed to improve the traditional SGD: based on parameter prediction, the historical training error is used to update the learning rate. Under the condition of given initial learning rate, experiments on classical data sets prove the effectiveness of the above algorithms, and the adaptive learning rate stochastic gradient descent algorithm can keep the convergence of the network. Training accuracy is relatively stable; Shorter training time; And improve the learning accuracy.

Music sentiment classification based on an optimized CNN-RF-QPSO model

PurposeMusic sentiment analysis helps to promote the diversification of music information retrieval methods. Traditional music emotion classification tasks suffer from high manual workload and low classification accuracy caused by difficulty in feature extraction and inaccurate manual determination of hyperparameter. In this paper, the authors propose an optimized convolution neural network-random forest (CNN-RF) model for music sentiment classification which is capable of optimizing the manually selected hyperparameters to improve the accuracy of music sentiment classification and reduce labor costs and human classification errors.Design/methodology/approachA CNN-RF music sentiment classification model is designed based on quantum particle swarm optimization (QPSO). First, the audio data are transformed into a Mel spectrogram, and feature extraction is conducted by a CNN. Second, the music features extracted are processed by RF algorithm to complete a preliminary emotion classification. Finally, to select the suitable hyperparameters for a CNN, the QPSO algorithm is adopted to extract the best hyperparameters and obtain the final classification results.FindingsThe model has gone through experimental validations and achieved a classification accuracy of 97 per cent for different sentiment categories with shortened training time. The proposed method with QPSO achieved 1.2 and 1.6 per cent higher accuracy than that with particle swarm optimization and genetic algorithm, respectively. The proposed model had great potential for music sentiment classification.Originality/valueThe dual contribution of this work comprises the proposed model which integrated two deep learning models and the introduction of a QPSO into model optimization. With these two innovations, the efficiency and accuracy of music emotion recognition and classification have been significantly improved.

An Analysis of Deep Learning Models in SSVEP-Based BCI: A Survey.

The brain-computer interface (BCI), which provides a new way for humans to directly communicate with robots without the involvement of the peripheral nervous system, has recently attracted much attention. Among all the BCI paradigms, BCIs based on steady-state visual evoked potentials (SSVEPs) have the highest information transfer rate (ITR) and the shortest training time. Meanwhile, deep learning has provided an effective and feasible solution for solving complex classification problems in many fields, and many researchers have started to apply deep learning to classify SSVEP signals. However, the designs of deep learning models vary drastically. There are many hyper-parameters that influence the performance of the model in an unpredictable way. This study surveyed 31 deep learning models (2011-2023) that were used to classify SSVEP signals and analyzed their design aspects including model input, model structure, performance measure, etc. Most of the studies that were surveyed in this paper were published in 2021 and 2022. This survey is an up-to-date design guide for researchers who are interested in using deep learning models to classify SSVEP signals.

An Efficient Optimization Technique for Training Deep Neural Networks

Deep learning is a sub-branch of artificial intelligence that acquires knowledge by training a neural network. It has many applications in the field of banking, automobile industry, agriculture, and healthcare industry. Deep learning has played a significant role in solving complex tasks related to computer vision, such as image classification, natural language processing, and object detection. On the other hand, optimizers also play an intrinsic role in training the deep learning model. Recent studies have proposed many deep learning models, such as VGG, ResNet, DenseNet, and ImageNet. In addition, there are many optimizers such as stochastic gradient descent (SGD), Adam, AdaDelta, Adabelief, and AdaMax. In this study, we have selected those models that require lower hardware requirements and shorter training times, which facilitates the overall training process. We have modified the Adam based optimizers and minimized the cyclic path. We have removed an additional hyper-parameter from RMSProp and observed that the optimizer works with various models. The learning rate is set to minimum and constant. The initial weights are updated after each epoch, which helps to improve the accuracy of the model. We also changed the position of the epsilon in the default Adam optimizer. By changing the position of the epsilon, it accumulates the updating process. We used various models with SGD, Adam, RMSProp, and the proposed optimization technique. The results indicate that the proposed method is effective in achieving the accuracy and works well with the state-of-the-art architectures.

Classification and identification of garment images based on deep learning

The clothing images on the Internet is growing rapidly, and there is an increasing demand for the clothing images’ intelligent classification. In this paper, Region-Based Fully Convolutional Networks (R-FCN) is introduced into the clothing image recognition. In the clothing image classification, because the network training time is long and the recognition rate of deformed clothing images is low, an improved framework HSR-FCN is proposed. The regional suggestion network and HyperNet network in R-FCN are integrated in the new framework, the learning approach of image features is changed in HSR-FCN, the higher accuracy can be achieved in a shorter training time. A spatial transformation network is introduced into the model, the input clothing image and feature map are spatially transformed and aligned, the feature learning is strengthened for multi-angle clothing and deformed clothing. The experimental results show that the improved HSR-FCN model is used to strengthen effectively the learning of deformed clothing images, and with a shorter training time, the average accuracy rate of the original network model R-FCN is increased by about 3%, it reachs 96.69%.