Multi-task Learning Neural Network Research Articles

Real-time equilibrium reconstruction by multi-task learning neural network based on HL-3 tokamak

Abstract A neural network model, EFITNN, has been developed capable of real-time magnetic equilibrium reconstruction based on HL-3 tokamak magnetic measurement signals. The model processes inputs from 68 channels of magnetic measurement data gathered from 1159 HL-3 experimental discharges, including plasmas current, loop voltage, and the poloidal magnetic fields measured by probes. The outputs of the model feature eight key plasmas parameters, alongside high-resolution ( 129 × 129 ) reconstructions of the toroidal current density J P and poloidal magnetic flux distributions Ψ rz . Moreover, the network’s architecture employs a multi-task learning structure, which enables the sharing of weights and mutual correction among different outputs, and leads to increase the model’s accuracy by up to 32%. The performance of EFITNN demonstrates remarkable consistency with the offline EFIT, achieving average R 2 = 0.941, 0.997 and 0.959 for eight plasmas parameters, Ψ rz and J P , respectively. The model’s robust generalization capabilities are particularly evident in its successful predictions of quasi-snowflake (QSF) divertor configurations and its adept handling of data from shot numbers or plasmas current intervals not previously encountered during training. Compared to numerical methods, EFITNN significantly enhances computational efficiency with average computation time ranging from 0.08 ms to 0.45 ms, indicating its potential utility in real-time isoflux control and plasmas profile management.

Multi-task learning neural network for monitoring and diagnosis of smart meters in power IoT systems

Multi-task learning neural network for monitoring and diagnosis of smart meters in power IoT systems

Deep learning-based upgoing and downgoing wavefield separation for vertical seismic profile data

The vertical seismic profile (VSP) considerably aids attenuation analysis and velocity calibration, enabling high-resolution seismic exploration. However, the imaging and interpretation of VSP data require pure upgoing and downgoing wavefields, and their separation is critical in processing and interpretation. Often, achieving high precision and efficient wavefield separation can be regarded as a challenging issue. To automate this process with higher accuracy, we develop a deep-learning high-precision intelligent separation method. First, to construct a physics-driven and data-generic training set, the well-log data from different geologic environments are used to simulate the upgoing and downgoing VSP data independently. Thereafter, we train a multitask-learning neural network to regress the network weights before and after separation to tune a mapping model, which is used to implement intelligent upgoing and downgoing wavefield separation. Validation data confirm the feasibility of our method. To further study the generalizability and robustness of the method, we apply the trained model to separate the modeled and real VSP data. By analyzing the separated profiles and frequency-wavenumber spectra, it is observed that the intelligent separation method is superior to conventional methods with a balance between accuracy and efficiency. It can be concluded that our method can separate upgoing and downgoing wavefields in VSP data with excellent performance, even in complex geologic environments. Our study indicates that the multitask-learning neural network combined with a physics-driven and data-generic training set is a new strategy to separate upgoing and downgoing wavefields and deserves to be popularized and applied.

Plane Cascade Aerodynamic Performance Prediction Based on Metric Learning for Multi-Output Gaussian Process Regression

Multi-output Gaussian process regression measures the similarity between samples based on Euclidean distance and assigns the same weight to each feature. However, there are significant differences in the aerodynamic performance of plane cascades composed of symmetric and asymmetric blade shapes, and there are also significant differences between the geometry of the plane cascades formed by different blade shapes and the experimental working conditions. There are large differences in geometric and working condition parameters in the features, which makes it difficult to accurately measure the similarity between different samples when there are fewer samples. For this problem, a metric learning for the multi-output Gaussian process regression method (ML_MOGPR) for aerodynamic performance prediction of the plane cascade is proposed. It shares parameters between multiple output Gaussian distributions during training and measures the similarity between input samples in a new embedding space to reduce bias and improve overall prediction accuracy. For the analysis of ML_MOGPR prediction results, the overall prediction accuracy is significantly improved compared with multi-output Gaussian process regression (MOGPR), backpropagation neural network (BPNN), and multi-task learning neural network (MTLNN). The experimental results show that ML_MOGPR is effective in predicting the performance of the plane cascade, and it can quickly and accurately make a preliminary estimate of the aerodynamic performance and meet the performance parameter estimation accuracy requirements in the early stage.

Heuristic optimisation of multi-task dynamic architecture neural network (DAN2)

This article proposes a novel method to optimise the Dynamic Architecture Neural Network (DAN2) adapted for a multi-task learning problem. The multi-task learning neural network adopts a multi-head and serial architecture with DAN2 layers acting as the basic subroutine. Adopting a dynamic architecture, the layers are added consecutively starting from a minimal initial structure. The optimisation method adopts an iterative heuristic scheme that sequentially optimises the shared layers and the task-specific layers until the solver converges to a small tolerance. Application of the method has demonstrated the applicability of the algorithm to simulated datasets. Comparable results to Artificial Neural Networks (ANNs) have been obtained in terms of accuracy and speed.

PPG-based blood pressure estimation can benefit from scalable multi-scale fusion neural networks and multi-task learning

The continuous measurement of blood pressure (BP) plays an important role in preventing cardiovascular diseases. However, common cuff-based devices with cumbersome operations are difficult to realize continuous BP monitoring. We propose a continuous BP estimation method based on the end-to-end multi-scale fusion and multi-task learning neural network (MSF-MTLNet). It can use the 8-second PPG segment and its derivatives as input and automatically learn their multi-scale features to improve BP accuracy. The proposed method implements feature extraction by adaptively fusing multiple convolution kernels with different sizes through the attention mechanism. Simultaneously, multi-task learning was introduced to discover task-relevant features. A total of 277,600 segments from 1825 patients in the University of California, Irvine database are used for experiments. They are divided into the independent training set, validation set, and test set according to the ratio of 8:1:1 in an inter-subject manner. The mean and standard deviation of estimation error for systolic blood pressure and diastolic blood pressure are 0.97 ± 8.87 mmHg and 0.55 ± 4.23 mmHg, respectively, which outperforms the current cutting-edge method. Our approach almost meets the Association for the Advancement of Medical Instrumentation standard. Considering the British Hypertension Society standards, it achieves an 'A' grade for DBP estimation and a 'B' grade for SBP estimation. Experimental results show that MSF-MTLNet can obtain promising precision in cuffless and continuous BP estimation.

Nuclear mass based on the multi-task learning neural network method

Nuclear mass based on the multi-task learning neural network method

A Multiobjective Cooperative Driving Framework Based on Evolutionary Algorithm and Multitask Learning

The development of connected and automated vehicle (CAV) techniques brings an upcoming revolution to traffic management. The control of CAVs in potential conflict areas such as on-ramps and intersections will be complex to traffic management when considering their deployment. There is still a lack of a general framework for dispatching CAVs in these bottlenecks, which is expected to ensure safety, traffic efficiency, and energy consumption in real time. This study aimed to fill the technique gap, and a comprehensive cooperative intelligent driving framework is put forward to study the problem, which can be used in both on-ramp and intersection scenarios. Based on a multi-objective evolutionary algorithm, CAVs are denoted as a sequence to be searched in solution space, while a multitask learning neural network with adaptive loss function is implemented for optimization target feedback to surrogate the simulation test procedure. The simulation results show that the proposed framework can get satisfying performance with low time and energy consumption. It can reduce time consumption by up to 16.51% for the on-ramp scenario and 9.8% for the intersection scenario, while reducing energy consumption by up to 16.39% and 11.39% for the two scenarios. Meanwhile, an analysis of computation time is carried out, illuminating the flexibility and controllability of the new strategy.

Topic Break Detection in Interview Dialogues Using Sentence Embedding of Utterance and Speech Intention Based on Multitask Neural Networks.

Currently, task-oriented dialogue systems that perform specific tasks based on dialogue are widely used. Moreover, research and development of non-task-oriented dialogue systems are also actively conducted. One of the problems with these systems is that it is difficult to switch topics naturally. In this study, we focus on interview dialogue systems. In an interview dialogue, the dialogue system can take the initiative as an interviewer. The main task of an interview dialogue system is to obtain information about the interviewee via dialogue and to assist this individual in understanding his or her personality and strengths. In order to accomplish this task, the system needs to be flexible and appropriate for detecting topic switching and topic breaks. Given that topic switching tends to be more ambiguous in interview dialogues than in task-oriented dialogues, existing topic modeling methods that determine topic breaks based only on relationships and similarities between words are likely to fail. In this study, we propose a method for detecting topic breaks in dialogue to achieve flexible topic switching in interview dialogue systems. The proposed method is based on multi-task learning neural network that uses embedded representations of sentences to understand the context of the text and utilizes the intention of an utterance as a feature. In multi-task learning, not only topic breaks but also the intention associated with the utterance and the speaker are targets of prediction. The results of our evaluation experiments show that using utterance intentions as features improves the accuracy of topic separation estimation compared to the baseline model.

Multi-task learning for digital rock segmentation and characteristic parameters computation

“Image-and-compute” is the key paradigm of digital rock physics (DRP), the two main processes are relatively independent. Deep learning algorithm has been well applied in the field of digital rock image processing and numerical computation, but the models of different tasks are lack of correlation, and there is still a large space to improve the accuracy and efficiency. Multi-task learning algorithm can integrate image processing and numerical computation technology into the same model and share effective information of different tasks, which has great significance to simplify the workflow of DRP and improve the computation accuracy of the model. A distributed multi-task learning neural network (DMTNN) is designed by simulating the working sequence of DRP, which can accomplish digital rock image segmentation, petrophysical and elastic parameters computation simultaneously. DMTNN adopts the strategy of series connection and parallel connection between tasks, the segmentation and petrophysical parameters computation tasks provide intuitive low-level features as auxiliary information to improve the accuracy of abstract high-level targets. Moreover, a dynamic weight strategy is applied to solve the problem of unbalanced convergence caused by mutual interference between tasks in the training process. In order to verify the effectiveness of the proposed method, six different open source sandstones are used to form the training-set and testing-set. The results show that the average pixel accuracy (PA) of DMTNN's segmentation task is 0.97, the R2-score of porosity, shear modulus and bulk modulus can reach 0.92, 0.74 and 0.79 respectively. Furthermore, another completely untrained digital rock data is selected as the cross-dataset, in order to verify the robustness of the proposed model. The results show DMTNN in cross-dataset: the PA of segmentation task can reach more than 0.95, and the relative error of the elastic parameters computation tasks are lower than the traditional CNN.

YolTrack: Multitask Learning Based Real-Time Multiobject Tracking and Segmentation for Autonomous Vehicles.

Modern autonomous vehicles are required to perform various visual perception tasks for scene construction and motion decision. The multiobject tracking and instance segmentation (MOTS) are the main tasks since they directly influence the steering and braking of the car. Implementing both tasks using a multitask learning neural network presents significant challenges in performance and complexity. Current work on MOTS devotes to improve the precision of the network with a two-stage tracking by detection model, which is difficult to satisfy the real-time requirement of autonomous vehicles. In this article, a real-time multitask network named YolTrack based on one-stage instance segmentation model is proposed to perform the MOTS task, achieving an inference speed of 29.5 frames per second (fps) with slight accuracy and precision drop. The YolTrack uses ShuffleNet V2 with feature pyramid network (FPN) as a backbone, from which two decoders are extended to generate instance segments and embedding vectors. Segmentation masks are used to improve the tracking performance by performing logic AND operation with feature maps, proving that foreground segmentation plays an important role in object tracking. The different scales of multiple tasks are balanced by the optimized geometric mean loss during the training phase. Experimental results on the KITTI MOTS data set show that YolTrack outperforms other state-of-the-art MOTS architectures in real-time aspect and is appropriate for deployment in autonomous vehicles.

A Bayesian Driver Agent Model for Autonomous Vehicles System Based on Knowledge-Aware and Real-Time Data

A key research area in autonomous driving is how to model the driver’s decision-making behavior, due to the fact it is significant for a self-driving vehicles considering their traffic safety and efficiency. However, the uncertain characteristics of vehicle and pedestrian trajectories affect urban roads, which poses severe challenges to the cognitive understanding and decision-making of autonomous vehicle systems in terms of accuracy and robustness. To overcome the abovementioned problems, this paper proposes a Bayesian driver agent (BDA) model which is a vision-based autonomous vehicle system with learning and inference methods inspired by human driver’s cognitive psychology. Different from the end-to-end learning method and traditional rule-based methods, our approach breaks the driving system up into a scene recognition module and a decision inference module. The perception module, which is based on a multi-task learning neural network (CNN), takes a driver’s-view image as its input and predicts the traffic scene’s feature values. The decision module based on dynamic Bayesian network (DBN) then makes an inferred decision using the traffic scene’s feature values. To explore the validity of the Bayesian driver agent model, we performed experiments on a driving simulation platform. The BDA model can extract the scene feature values effectively and predict the probability distribution of the human driver’s decision-making process accurately based on inference. We take the lane changing scenario as an example to verify the model, the intraclass correlation coefficient (ICC) correlation between the BDA model and human driver’s decision process reached 0.984. This work suggests a research in scene perception and autonomous decision-making that may apply to autonomous vehicle system.

Reinforcement Learning and Deep Learning Based Lateral Control for Autonomous Driving [Application Notes

This paper investigates the vision-based autonomous driving with deep learning and reinforcement learning methods. Different from the end-to-end learning method, our method breaks the vision-based lateral control system down into a perception module and a control module. The perception module which is based on a multi-task learning neural network first takes a driver-view image as its input and predicts the track features. The control module which is based on reinforcement learning then makes a control decision based on these features. In order to improve the data efficiency, we propose visual TORCS (VTORCS), a deep reinforcement learning environment which is based on the open racing car simulator (TORCS). By means of the provided functions, one can train an agent with the input of an image or various physical sensor measurement, or evaluate the perception algorithm on this simulator. The trained reinforcement learning controller outperforms the linear quadratic regulator (LQR) controller and model predictive control (MPC) controller on different tracks. The experiments demonstrate that the perception module shows promising performance and the controller is capable of controlling the vehicle drive well along the track center with visual input.

Deep Convolutional Neural Network with Optical Flow for Facial Micro-Expression Recognition

Micro-expression is a kind of brief facial movements which could not be controlled by the nervous system. Micro-expression indicates that a person is hiding his true emotion consciously. Micro-expression recognition has various potential applications in public security and clinical medicine. Researches are focused on the automatic micro-expression recognition, because it is hard to recognize the micro-expression by people themselves. This research proposed a novel algorithm for automatic micro-expression recognition which combined a deep multi-task convolutional network for detecting the facial landmarks and a fused deep convolutional network for estimating the optical flow features of the micro-expression. First, the deep multi-task convolutional network is employed to detect facial landmarks with the manifold-related tasks for dividing the facial region. Furthermore, a fused convolutional network is applied for extracting the optical flow features from the facial regions which contain the muscle changes when the micro-expression appears. Because each video clip has many frames, the original optical flow features of the whole video clip will have high number of dimensions and redundant information. This research revises the optical flow features for reducing the redundant dimensions. Finally, a revised optical flow feature is applied for refining the information of the features and a support vector machine classifier is adopted for recognizing the micro-expression. The main contribution of work is combining the deep multi-task learning neural network and the fusion optical flow network for micro-expression recognition and revising the optical flow features for reducing the redundant dimensions. The results of experiments on two spontaneous micro-expression databases prove that our method achieved competitive performance in micro-expression recognition.