Encoder Network Research Articles

Multi parallel U-net encoder network for effective polyp image segmentation

Automating polyps' detection and segmentation is very helpful in achieving early detection and potentially effective treatment of colon cancer. Because of the constant rise in processing power at a reasonable cost and the availability of training data, deep learning more than ever is becoming very powerful and increasingly widespread in many domains, including computer vision and medical image processing and segmentation. In this paper, we propose a novel method for polyp image segmentation using convolutional neural networks (CNNs) and utilizing multi-parallel U-Net encoder architecture. The proposed method uses multiple encoders that utilize pre-trained CNNs to enrich the extracted features. Skip connections from these parallel encoders are concatenated and propagated to the decoder in partially and fully meshed fashions allowing for data from shallow to deeper layers in the network to be exchanged and vice versa. We trained and tested the method on five publicly available datasets: Kvasir, CVC-Clinic DB, CVC-Colon DB, CVC-T, and ETIS-Larib, which are well-known benchmark datasets for polyp image segmentation. We tested the performance of the method on these five datasets using multiple testing scenarios including the effect of using attention, vision transformers, and multiple decoders. Experimental results showed that the proposed method outperforms other state-of-the-art methods on two of the used benchmark datasets and ranked second on a third.

Subject-Independent EEG Classification of Motor Imagery Based on Dual-Branch Feature Fusion.

A brain computer interface (BCI) system helps people with motor dysfunction interact with the external environment. With the advancement of technology, BCI systems have been applied in practice, but their practicability and usability are still greatly challenged. A large amount of calibration time is often required before BCI systems are used, which can consume the patient's energy and easily lead to anxiety. This paper proposes a novel motion-assisted method based on a novel dual-branch multiscale auto encoder network (MSAENet) to decode human brain motion imagery intentions, while introducing a central loss function to compensate for the shortcomings of traditional classifiers that only consider inter-class differences and ignore intra-class coupling. The effectiveness of the method is validated on three datasets, namely BCIIV2a, SMR-BCI and OpenBMI, to achieve zero calibration of the MI-BCI system. The results show that our proposed network displays good results on all three datasets. In the case of subject-independence, the MSAENet outperformed the other four comparison methods on the BCIIV2a and SMR-BCI datasets, while achieving F1_score values as high as 69.34% on the OpenBMI dataset. Our method maintains better classification accuracy with a small number of parameters and short prediction times, and the method achieves zero calibration of the MI-BCI system.

Text Classification of Mixed Model Based on Deep Learning

At present, deep learning has been widely used many fields, but the research on text classification is still relatively few. This paper makes full use of the good learning characteristics of deep learning, proposes a hybrid model based on deep learning, and designs a text classifier based on the hybrid model. This hybrid model uses two common deep learning models, sparse automatic encoder and deep confidence network, to mix. The hybrid model is mainly composed of three parts, the first two layers are constructed by sparse automatic encoder, the middle layer is a three-layer depth Convolutional Neural Network (CNN), and finally Softmax regression is used as the classification layer. In order to test the classification performance of the classifier based on deep learning hybrid model, relevant experiments were conducted on English data set 20Newsgroup and Chinese data set Fudan University Chinese Corpus. In the English text classification experiment, the classifier based on deep learning hybrid model is used to classify, and a high classification accuracy rate is obtained. In order to further verify the superiority of its performance, a comparative experiment with naive Bayes classifier, K-Nearest Neighbor (KNN) classifier and Support Vector Machine (SVM) classifier demonstrates that the classification effect of the classifier based on deep learning hybrid model is better than that of naive Bayes classifier, KNN classifier and support vector machine classifier. In the experiment of Chinese text classification, the Chinese corpus of Fudan University is tested, and a good classification effect is obtained. The influence of different parameter settings on the classification accuracy is discussed.

Sports Risk Prediction Model Based on Automatic Encoder and Convolutional Neural Network

In view of the limitations of traditional statistical methods in dealing with multifactor and nonlinear data and the inadequacy of classical machine learning algorithms in dealing with and predicting data with high dimensions and large sample sizes, this paper proposes an operational risk prediction model based on an automatic encoder and convolutional neural networks. First, we use an automatic encoder to extract features of motion risk factors and obtain feature components that can highly represent risk. Secondly, based on the causal relationship between sports risk and risk characteristics, a convolutional neural network with a dual convolution layer and dual pooling layer topology is constructed. Finally, the sports risk prediction model is established by combining the auto-coded feature components with the topology of the convolutional neural network. Compared with other algorithms, the proposed method can effectively analyze and extract risk characteristics and has a high prediction accuracy. At the same time, it promotes the integration of sports science and computer science and provides a basis for the application of machine learning in the field of sports risk prediction.

Simultaneous dose distribution and fluence prediction for nasopharyngeal carcinoma IMRT

BackgroundCurrent intensity-modulated radiation therapy (IMRT) treatment planning is still a manual and time/resource consuming task, knowledge-based planning methods with appropriate predictions have been shown to enhance the plan quality consistency and improve planning efficiency. This study aims to develop a novel prediction framework to simultaneously predict dose distribution and fluence for nasopharyngeal carcinoma treated with IMRT, the predicted dose information and fluence can be used as the dose objectives and initial solution for an automatic IMRT plan optimization scheme, respectively.MethodsWe proposed a shared encoder network to simultaneously generate dose distribution and fluence maps. The same inputs (three-dimensional contours and CT images) were used for both dose distribution and fluence prediction. The model was trained with datasets of 340 nasopharyngeal carcinoma patients (260 cases for training, 40 cases for validation, 40 cases for testing) treated with nine-beam IMRT. The predicted fluence was then imported back to treatment planning system to generate the final deliverable plan. Predicted fluence accuracy was quantitatively evaluated within projected planning target volumes in beams-eye-view with 5 mm margin. The comparison between predicted doses, predicted fluence generated doses and ground truth doses were also conducted inside patient body.ResultsThe proposed network successfully predicted similar dose distribution and fluence maps compared with ground truth. The quantitative evaluation showed that the pixel-based mean absolute error between predicted fluence and ground truth fluence was 0.53% ± 0.13%. The structural similarity index also showed high fluence similarity with values of 0.96 ± 0.02. Meanwhile, the difference in the clinical dose indices for most structures between predicted dose, predicted fluence generated dose and ground truth dose were less than 1 Gy. As a comparison, the predicted dose achieved better target dose coverage and dose hot spot than predicted fluence generated dose compared with ground truth dose.ConclusionWe proposed an approach to predict 3D dose distribution and fluence maps simultaneously for nasopharyngeal carcinoma patients. Hence, the proposed method can be potentially integrated in a fast automatic plan generation scheme by using predicted dose as dose objectives and predicted fluence as a warm start.

Visual Question Generation Answering (VQG-VQA) using Machine Learning Models

Presented automated visual question-answer system generates graphics-based question-answer pairs. The system consists of the Visual Query Generation (VQG) and Visual Question Answer (VQA) modules. VQG generates questions based on visual cues, and VQA provides matching answers to the VQG modules. VQG system generates questions using LSTM and VGG19 model, training parameters, and predicting words with the highest probability for output. VQA uses VGG-19 convolutional neural network for image encoding, embedding, and multilayer perceptron for high-quality responses. The proposed system reduces the need for human annotation and thus supports the traditional education sector by significantly reducing the human intervention required to generate text queries. The system can be used in interactive interfaces to help young children learn.

CoordFill: Efficient High-Resolution Image Inpainting via Parameterized Coordinate Querying

Image inpainting aims to fill the missing hole of the input. It is hard to solve this task efficiently when facing high-resolution images due to two reasons: (1) Large reception field needs to be handled for high-resolution image inpainting. (2) The general encoder and decoder network synthesizes many background pixels synchronously due to the form of the image matrix. In this paper, we try to break the above limitations for the first time thanks to the recent development of continuous implicit representation. In detail, we down-sample and encode the degraded image to produce the spatial-adaptive parameters for each spatial patch via an attentional Fast Fourier Convolution (FFC)-based parameter generation network. Then, we take these parameters as the weights and biases of a series of multi-layer perceptron (MLP), where the input is the encoded continuous coordinates and the output is the synthesized color value. Thanks to the proposed structure, we only encode the high-resolution image in a relatively low resolution for larger reception field capturing. Then, the continuous position encoding will be helpful to synthesize the photo-realistic high-frequency textures by re-sampling the coordinate in a higher resolution. Also, our framework enables us to query the coordinates of missing pixels only in parallel, yielding a more efficient solution than the previous methods. Experiments show that the proposed method achieves real-time performance on the 2048X2048 images using a single GTX 2080 Ti GPU and can handle 4096X4096 images, with much better performance than existing state-of-the-art methods visually and numerically. The code is available at: https://github.com/NiFangBaAGe/CoordFill.

Pointerformer: Deep Reinforced Multi-Pointer Transformer for the Traveling Salesman Problem

Traveling Salesman Problem (TSP), as a classic routing optimization problem originally arising in the domain of transportation and logistics, has become a critical task in broader domains, such as manufacturing and biology. Recently, Deep Reinforcement Learning (DRL) has been increasingly employed to solve TSP due to its high inference efficiency. Nevertheless, most of existing end-to-end DRL algorithms only perform well on small TSP instances and can hardly generalize to large scale because of the drastically soaring memory consumption and computation time along with the enlarging problem scale. In this paper, we propose a novel end-to-end DRL approach, referred to as Pointerformer, based on multi-pointer Transformer. Particularly, Pointerformer adopts both reversible residual network in the encoder and multi-pointer network in the decoder to effectively contain memory consumption of the encoder-decoder architecture. To further improve the performance of TSP solutions, Pointerformer employs a feature augmentation method to explore the symmetries of TSP at both training and inference stages as well as an enhanced context embedding approach to include more comprehensive context information in the query. Extensive experiments on a randomly generated benchmark and a public benchmark have shown that, while achieving comparative results on most small-scale TSP instances as state-of-the-art DRL approaches do, Pointerformer can also well generalize to large-scale TSPs.

RWEN-TTS: Relation-Aware Word Encoding Network for Natural Text-to-Speech Synthesis

With the advent of deep learning, a huge number of text-to-speech (TTS) models which produce human-like speech have emerged. Recently, by introducing syntactic and semantic information w.r.t the input text, various approaches have been proposed to enrich the naturalness and expressiveness of TTS models. Although these strategies showed impressive results, they still have some limitations in utilizing language information. First, most approaches only use graph networks to utilize syntactic and semantic information without considering linguistic features. Second, most previous works do not explicitly consider adjacent words when encoding syntactic and semantic information, even though it is obvious that adjacent words are usually meaningful when encoding the current word. To address these issues, we propose Relation-aware Word Encoding Network (RWEN), which effectively allows syntactic and semantic information based on two modules (i.e., Semantic-level Relation Encoding and Adjacent Word Relation Encoding). Experimental results show substantial improvements compared to previous works.

Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network

Contrastive loss has significantly improved performance in supervised classification tasks by using a multi-viewed framework that leverages augmentation and label information. The augmentation enables contrast with another view of a single image but enlarges training time and memory usage. To exploit the strength of multi-views while avoiding the high computation cost, we introduce a multi-exit architecture that outputs multiple features of a single image in a single-viewed framework. To this end, we propose Self-Contrastive (SelfCon) learning, which self-contrasts within multiple outputs from the different levels of a single network. The multi-exit architecture efficiently replaces multi-augmented images and leverages various information from different layers of a network. We demonstrate that SelfCon learning improves the classification performance of the encoder network, and empirically analyze its advantages in terms of the single-view and the sub-network. Furthermore, we provide theoretical evidence of the performance increase based on the mutual information bound. For ImageNet classification on ResNet-50, SelfCon improves accuracy by +0.6% with 59% memory and 48% time of Supervised Contrastive learning, and a simple ensemble of multi-exit outputs boosts performance up to +1.5%. Our code is available at https://github.com/raymin0223/self-contrastive-learning.

Skin Lesion Intelligent Diagnosis in Edge Computing Networks: An FCL Approach

In recent years, automatic skin lesion diagnosis methods based on artificial intelligence have achieved great success. However, the lack of labeled data, visual similarity between skin diseases, and restriction on private data sharing remain the major challenges in skin lesion diagnosis. In this article, first, we propose a federated contrastive learning framework to break down data silos and enhance the generalizability of diagnostic model to unseen data. Subsequently, by combining data features from different participated nodes, the proposed framework can improve the performance of contrastive training. To extract discriminative features during on-device training, we propose a contrastive learning based intelligent skin lesion diagnosis scheme in edge computing networks. Specifically, a contrastive learning based dual encoder network is designed to overcome training sample scarcity by fully leveraging unlabeled samples for performance improvement. Meanwhile, we devise a maximum mean discrepancy based supervised contrastive loss function, which can efficiently explore complex intra-class and inter-class variances of samples. Finally, the diagnosis simulations demonstrate that compared with existing methods, our proposed scheme can achieve superior accuracy in both on-device training and distributed training scenarios.

Bidirectional feature matching based on deep pairwise contrastive learning for multiparametric MRI image synthesis

Objective. Multi-parametric MR image synthesis is an effective approach for several clinical applications where specific modalities may be unavailable to reach a diagnosis. While technical and practical conditions limit the acquisition of new modalities for a patient, multimodal image synthesis combines multiple modalities to synthesize the desired modality. Approach. In this paper, we propose a new multi-parametric magnetic resonance imaging (MRI) synthesis model, which generates the target MRI modality from two other available modalities, in pathological MR images. We first adopt a contrastive learning approach that trains an encoder network to extract a suitable feature representation of the target space. Secondly, we build a synthesis network that generates the target image from a common feature space that approximately matches the contrastive learned space of the target modality. We incorporate a bidirectional feature learning strategy that learns a multimodal feature matching function, in two opposite directions, to transform the augmented multichannel input in the learned target space. Overall, our training synthesis loss is expressed as the combination of the reconstruction loss and a bidirectional triplet loss, using a pair of features. Main results. Compared to other state-of-the-art methods, the proposed model achieved an average improvement rate of 3.9% and 3.6% on the IXI and BraTS’18 datasets respectively. On the tumor BraTS’18 dataset, our model records the highest Dice score of 0.793(0.04) for preserving the synthesized tumor regions in the segmented images. Significance. Validation of the proposed model on two public datasets confirms the efficiency of the model to generate different MR contrasts, and preserve tumor areas in the synthesized images. In addition, the model is flexible to generate head and neck CT image from MR acquisitions. In future work, we plan to validate the model using interventional iMRI contrasts for MR-guided neurosurgery applications, and also for radiotherapy applications. Clinical measurements will be collected during surgery to evaluate the model’s performance.

Zero-shot style transfer for gesture animation driven by text and speech using adversarial disentanglement of multimodal style encoding.

Modeling virtual agents with behavior style is one factor for personalizing human-agent interaction. We propose an efficient yet effective machine learning approach to synthesize gestures driven by prosodic features and text in the style of different speakers including those unseen during training. Our model performs zero-shot multimodal style transfer driven by multimodal data from the PATS database containing videos of various speakers. We view style as being pervasive; while speaking, it colors the communicative behaviors expressivity while speech content is carried by multimodal signals and text. This disentanglement scheme of content and style allows us to directly infer the style embedding even of a speaker whose data are not part of the training phase, without requiring any further training or fine-tuning. The first goal of our model is to generate the gestures of a source speaker based on the content of two input modalities-Mel spectrogram and text semantics. The second goal is to condition the source speaker's predicted gestures on the multimodal behavior style embedding of a target speaker. The third goal is to allow zero-shot style transfer of speakers unseen during training without re-training the model. Our system consists of two main components: (1) a speaker style encoder network that learns to generate a fixed-dimensional speaker embedding style from a target speaker multimodal data (mel-spectrogram, pose, and text) and (2) a sequence-to-sequence synthesis network that synthesizes gestures based on the content of the input modalities-text and mel-spectrogram-of a source speaker and conditioned on the speaker style embedding. We evaluate that our model is able to synthesize gestures of a source speaker given the two input modalities and transfer the knowledge of target speaker style variability learned by the speaker style encoder to the gesture generation task in a zero-shot setup, indicating that the model has learned a high-quality speaker representation. We conduct objective and subjective evaluations to validate our approach and compare it with baselines.

MR image harmonization based on cross‐center style transfer using deep generative adversarial network

AbstractBackgroundDue to heterogeneities in protocols such as magnetic field strength, coil channels, and scanner manufacturers, data shift problems occur. This causes a serious performance drop in predictions tested to other hospitals. We proposed a multi center‐cross transfer network aimed to adapt institutions with minimal loss of predictions for Alzheimer disease.MethodThree datasets collected from hospitals with different manufacturers and protocols were included in this study (Table1). We used a deep generative adversarial network (GAN) consisting of 4 models: The generator learns to generate images by inputting a source image and a style code. The style encoder and mapping network extracts the style code of images and transforms into style spaces of multiple domains. The discriminator distinguishes between real and fake images from multiple domains. The integrated model was adjusted while training based on four losses: adversarial loss, cycle‐consistency loss, style‐reconstruction loss, and style‐diversification loss. To resolve the data imbalance problem, 1200 slices were randomly extracted per institution. For validation, we evaluated how well the proposed deep learning model transforms images while maintaining image quality and variation. Next, classification was performed using a convolutional neural network model (Resnet 50) to check how successfully transfers styles of an institution to another one.ResultWe quantitatively evaluated transfering styles between the desired source and reference institution (Figures 1 and 2). Through interpolation and consistency loss coefficient differences, the source to reference transformation process can be checked continuously (Figure3). The proposed model performed well with SSIM of 0.897, LPIPS of 0.862, which are metrics for assessing quality and variation of image transfering model. To predict harmonization ability, we compared classification performance between each pair of datasets before and after style transfering, respectively. The accuracies for each pair shows dramatic decrese in performance of discriminating different centers: ADNI‐SMC (91.20 to 64.35), ADNI‐GMC (96.52 to 66.66) and GMC‐SMC (87.84 to 57.89).ConclusionWe proposed a multi center‐cross transfer network that freely converts protocol at any desired institution. This work will be able to predict Alzheimer’s biomarkers robustly by harmonizing all data from multiple hospitals, while preserving disease properties encoded in the dataset.

Visual Image Decoding of Brain Activities Using a Dual Attention Hierarchical Latent Generative Network With Multiscale Feature Fusion

Reconstructing visual stimulus from human brain activity measured with functional magnetic resonance imaging (fMRI) is a challenging decoding task for revealing the visual system. Recent deep learning approaches commonly neglect the relationship between hierarchical image features and different regions of the visual cortex, and fail to use global and local image features in reconstructing visual stimulus. To address these issues, in this paper, a novel neural decoding framework is proposed by using a dual attention hierarchical latent generative network with multi-scale feature fusion (DA-HLGN-MSFF) method. Specifically, the fMRI data is firstly encoded to hierarchical features of our image encoder network which employs a multi-kernel convolution block to extract the multi-scale spatial information of images. In order to reconstruct the perceived images and further improve the performance of our generator network, a dual attention block based on channel-spatial attention mechanism is then proposed to exploit the inter-channel relationships and spatial long-range dependencies of features. Moreover, a multi-scale feature fusion block is finally adopted to aggregate the global and local information of features at different scales and synthesize the final reconstructed images in the generator network. Competitive experimental results on two public fMRI datasets demonstrate that our method is able to achieve promising reconstructing performance compared with the state-of-the-art methods. The codes of our proposed DA-HLGN-MSFF method will be open access on https://github.com/ljbuaa/HLDAGN.

Marbling-Net: A Novel Intelligent Framework for Pork Marbling Segmentation Using Images from Smartphones.

Marbling characteristics are important traits for the genetic improvement of pork quality. Accurate marbling segmentation is the prerequisite for the quantification of these traits. However, the marbling targets are small and thin with dissimilar sizes and shapes and scattered in pork, complicating the segmentation task. Here, we proposed a deep learning-based pipeline, a shallow context encoder network (Marbling-Net) with the usage of patch-based training strategy and image up-sampling to accurately segment marbling regions from images of pork longissimus dorsi (LD) collected by smartphones. A total of 173 images of pork LD were acquired from different pigs and released as a pixel-wise annotation marbling dataset, the pork marbling dataset 2023 (PMD2023). The proposed pipeline achieved an IoU of 76.8%, a precision of 87.8%, a recall of 86.0%, and an F1-score of 86.9% on PMD2023, outperforming the state-of-art counterparts. The marbling ratios in 100 images of pork LD are highly correlated with marbling scores and intramuscular fat content measured by the spectrometer method (R2 = 0.884 and 0.733, respectively), demonstrating the reliability of our method. The trained model could be deployed in mobile platforms to accurately quantify pork marbling characteristics, benefiting the pork quality breeding and meat industry.

Transfer Ranking in Finance: Applications to Cross-Sectional Momentum with Data Scarcity

Modern cross-sectional strategies incorporating sophisticated neural architectures outperform their traditional counterparts when applied to mature assets with long histories. However, deploying them on instruments with limited samples generally produces over-fitted models with degraded performance. In this paper, we introduce Fused Encoder Networks -- a hybrid parameter-sharing transfer ranking model which fuses information extracted using an encoder-attention module from a source dataset with a similar but separate module operating on a smaller target dataset of interest. This mitigates the issue of models with poor generalisability. Additionally, the self-attention mechanism enables interactions among instruments to be accounted for, both at the loss-level during model training and at inference time. Focusing on momentum applied to the top ten cryptocurrencies by market capitalisation as a demonstrative use-case, our model outperforms state-of-the-art benchmarks on most measures and significantly boosts the Sharpe ratio. It continues outperforming baselines even after accounting for the high transaction costs associated with trading cryptocurrencies.

Infrared and visible image fusion based on a two-stage class conditioned auto-encoder network

Existing auto-encoder based infrared and visible image fusion methods typically utilize a shared encoder to extract features from different modalities and adopt a handcrafted fusion strategy to fuse the extracted features into intermediate representation before the decoder part. In this paper, we present a novel two-stage class conditioned auto-encoder framework for high-quality multispectral fusion tasks. In the first training stage, we introduce a class embedding sub-branch to the encoder network for modeling the characteristics of different modalities and adaptively scaling the intermediate features based on the input modality. Moreover, we design a cross-transfer residual block to promote the content and texture information flow in the encoder for generating more representative features. In the second training stage, we insert a learnable fusion module between the pre-trained class conditioned encoder and decoder parts to replace the handcrafted fusion strategy. Specific intensity and gradient loss functions are utilized to tune the model for the fusion of distinctive deep features in a data-driven manner. With the important designs including the class conditioned auto-encoder and the two-stage training strategy, our proposed TS-ClassFuse can better preserve distinctive information/features from the source images and decrease the training difficulty for simultaneously extracting informative features and determining the optimal fusion scheme. Experimental results verify the effectiveness of our method in terms of both qualitative and quantitative evaluations.

CA-Net: Context Aggregation Network for Nuclei Classification in Histopathology Image

Accurately classifying nuclei in histopathology images is essential for cancer diagnosis and prognosis. However, due to the touching nuclei, nucleus shape variation, background complexity, and image artifacts, end-to-end nucleus classification is still difficult and challenging. In this manuscript, we propose a context aggregation network (CA-Net) for nuclei classification by fusing global contextual information which is critical for classifying nuclei in histopathology images. Specifically, we propose a multi-level semantic supervision (MSS) module focusing on extracting multi-scale context information by varying three different kernel sizes, and dynamically aggregating the context information from high to low level. Furthermore, we employ the GPG and SAPF modules in encoder and decoder networks to exact and aggregate global context information. Finally, the proposed network is verified on a mainstream nuclei classification image datasets (PanNuke) and achieves an improved global accuracy of 0.816. Our proposed MSS module can be easily transferred into any UNet-liked architecture as a deep supervision mechanism.

A Faster Dynamic Feature Extractor and Its Application to Industrial Quality Prediction

The unsupervised dynamic models have been applied to various tasks in the process industry due to their excellent ability to represent the process dynamics. The recurrent-network-based dynamic feature extractor is a typical unsupervised dynamic model which extracts the dynamic data features using a recurrent encoder network. However, the recurrent-network-based dynamic feature extractor has low computational efficiency due to its recurrent nature, which prevents the model from being used for large-scale data sets. To improve computational efficiency, a new dynamic feature extractor called TempoATTNE-DFE is proposed in this work. In TempoATTNE-DFE, a new encoder structure is developed, which can be implemented in parallel for data sequences. Meanwhile, a kind of attention mechanism is proposed to extract the dynamic features within the input sequence. The proposed TempoATTNE-DFE can achieve higher computational efficiency in offline training and online inference. To evaluate the effectiveness of TempoATTNE-DFE, it is applied to the quality prediction task and validated with a numerical example and two real-world industrial processes. The application results demonstrate that TempoATTNE-DFE can achieve better prediction performance compared to other state-of-the-art methods. In addition, compared with the recurrent-network-based dynamic feature extractor, TempoATTNE-DFE gains <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$ 1.29\times$</tex-math></inline-formula> speedup in training and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$ 2.45\times$</tex-math></inline-formula> speedup in inference on the blast furnace data set.