Showing papers on "Unsupervised learning published in 2023"

Machine learning based models for defect detection in composites inspected by Barker coded thermography: A qualitative analysis

Abstract: Machine learning and artificial intelligence have evolved as enablers for automation in various industrial applications. Barker-coded thermography is an active thermal non-destructive testing technique for examining subsurface features in industrial components. This article introduces supervised and unsupervised machine learning models for automatic defect detection in composite specimens inspected by the barker-coded stimulus. This work provides supervised and unsupervised machine learning methods to detect defects in composite specimens examined with a barker automatically coded stimulus. The suggested technology is tested using a carbon fiber-reinforced polymer sample with synthetically reproduced flat bottom hole flaws. The one-class Support vector machine is chosen for the unsupervised class of operation, whereas the supervised technique modifies the traditional Support Vector Machine (SVM). The qualitative comparison suggests that the unsupervised approach presents a less than 1% marginal difference in defect detection from its supervised counterpart.

Prototypical Contrast and Reverse Prediction: Unsupervised Skeleton Based Action Recognition

Abstract: We focus on unsupervised representation learning for skeleton based action recognition. Existing unsupervised approaches usually learn action representations by motion prediction but they lack the ability to fully learn inherent semantic similarity. In this paper, we propose a novel framework named Prototypical Contrast and Reverse Prediction (PCRP) to address this challenge. Different from plain motion prediction, PCRP performs reverse motion prediction based on encoder-decoder structure to extract more discriminative temporal pattern, and derives action prototypes by clustering to explore the inherent action similarity within the action encoding. Specifically, we regard action prototypes as latent variables and formulate PCRP as an expectation-maximization (EM) task. PCRP iteratively runs (1) E-step as to determine the distribution of action prototypes by clustering action encoding from the encoder while estimating concentration around prototypes, and (2) M-step as optimizing the model by minimizing the proposed ProtoMAE loss, which helps simultaneously pull the action encoding closer to its assigned prototype by contrastive learning and perform reverse motion prediction task. Besides, the sorting can also serve as a temporal task similar as reverse prediction in the proposed framework. Extensive experiments on N-UCLA, NTU 60, and NTU 120 dataset present that PCRP outperforms main stream unsupervised methods and even achieves superior performance over many supervised methods. The codes are available at: https://github.com/LZUSIAT/PCRP .

Self-supervised Learning: A Succinct Review

Abstract: Machine learning has made significant advances in the field of image processing. The foundation of this success is supervised learning, which necessitates annotated labels generated by humans and hence learns from labelled data, whereas unsupervised learning learns from unlabeled data. Self-supervised learning (SSL) is a type of un-supervised learning that helps in the performance of downstream computer vision tasks such as object detection, image comprehension, image segmentation, and so on. It can develop generic artificial intelligence systems at a low cost using unstructured and unlabeled data. The authors of this review article have presented detailed literature on self-supervised learning as well as its applications in different domains. The primary goal of this review article is to demonstrate how images learn from their visual features using self-supervised approaches. The authors have also discussed various terms used in self-supervised learning as well as different types of learning, such as contrastive learning, transfer learning, and so on. This review article describes in detail the pipeline of self-supervised learning, including its two main phases: pretext and downstream tasks. The authors have shed light on various challenges encountered while working on self-supervised learning at the end of the article.

A Comparative Analysis of Supervised and Unsupervised Models for Detecting Attacks on the Intrusion Detection Systems

Abstract: Intrusion Detection Systems are expected to detect and prevent malicious activities in a network, such as a smart grid. However, they are the main systems targeted by cyber-attacks. A number of approaches have been proposed to classify and detect these attacks, including supervised machine learning. However, these models require large labeled datasets for training and testing. Therefore, this paper compares the performance of supervised and unsupervised learning models in detecting cyber-attacks. The benchmark of CICDDOS 2019 was used to train, test, and validate the models. The supervised models are Gaussian Naïve Bayes, Classification and Regression Decision Tree, Logistic Regression, C-Support Vector Machine, Light Gradient Boosting, and Alex Neural Network. The unsupervised models are Principal Component Analysis, K-means, and Variational Autoencoder. The performance comparison is made in terms of accuracy, probability of detection, probability of misdetection, probability of false alarm, processing time, prediction time, training time per sample, and memory size. The results show that the Alex Neural Network model outperforms the other supervised models, while the Variational Autoencoder model has the best results compared to unsupervised models.

Unsupervised and supervised learning of interacting topological phases from single-particle correlation functions

Abstract: The recent advances in machine learning algorithms have boosted the application of these techniques to the field of condensed matter physics, in order e.g. to classify the phases of matter at equilibrium or to predict the real-time dynamics of a large class of physical models. Typically in these works, a machine learning algorithm is trained and tested on data coming from the same physical model. Here we demonstrate that unsupervised and supervised machine learning techniques are able to predict phases of a non-exactly solvable model when trained on data of a solvable model. In particular, we employ a training set made by single-particle correlation functions of a non-interacting quantum wire and by using principal component analysis, k-means clustering, and convolutional neural networks we reconstruct the phase diagram of an interacting superconductor. We show that both the principal component analysis and the convolutional neural networks trained on the data of the non-interacting model can identify the topological phases of the interacting model with a high degree of accuracy. Our findings indicate that non-trivial phases of matter emerging from the presence of interactions can be identified by means of unsupervised and supervised techniques applied to data of non-interacting systems.

Leakage detection in natural gas pipeline based on unsupervised learning and stress perception

Abstract: Natural gas pipeline leakage can cause serious financial losses to natural gas transportation and pose accidents to the environmental safety. Currently-used supervised learning methods heavily rely on sufficient pipeline failure historical data for their training. Therefore, we propose a novel detection approach based on unsupervised learning and stress perception for determining the leakage situation in pipelines. In this study, pipeline stress signals are first acquired based on residual magnetic effect. The relationship between residual magnetic and stress is built using improved sparrow search algorithm (ISSA) and extreme learning machine (ELM). Then, the Wasserstein generative adversarial network with gradient penalty (WGAN-GP) is deployed to learn suitable features from the stress signals under the pipeline normal condition, generating high-quality stress data features. Finally, the generated stress features are supplied to the Bayesian Gaussian mixture model (BGMM). And the weighted logarithm probability (WLP) is used as the health indicator for examining pipeline status. The results demonstrate that the relative error of residual magnetic stress model is controlled within 3 %, and the WLP value of fault samples is smaller than − 100, so that the proposed method can discriminate the normal and leak conditions as well as the risk and severity of leakage. This study provides a theoretical basis and new perspective for pipeline leakage detection.

An Efficient Unsupervised Learning Approach for Detecting Anomaly in Cloud

Abstract: The Cloud system shows its growing functionalities in various industrial applications. The safety towards data transfer seems to be a threat where Network Intrusion Detection System (NIDS) is measured as an essential element to fulfill security. Recently, Machine Learning (ML) approaches have been used for the construction of intellectual IDS. Most IDS are based on ML techniques either as unsupervised or supervised. In supervised learning, NIDS is based on labeled data where it reduces the efficiency of the reduced model to identify attack patterns. Similarly, the unsupervised model fails to provide a satisfactory outcome. Hence, to boost the functionality of unsupervised learning, an effectual auto-encoder is applied for feature selection to select good features. Finally, the Naïve Bayes classifier is used for classification purposes. This approach exposes the finest generalization ability to train the data. The unlabelled data is also used for adoption towards data analysis. Here, redundant and noisy samples over the dataset are eliminated. To validate the robustness and efficiency of NIDS, the anticipated model is tested over the NSL-KDD dataset. The experimental outcomes demonstrate that the anticipated approach attains superior accuracy with 93%, which is higher compared to J48, AB tree, Random Forest (RF), Regression Tree (RT), Multi-Layer Perceptrons (MLP), Support Vector Machine (SVM), and Fuzzy. Similarly, False Alarm Rate (FAR) and True Positive Rate (TPR) of Naive Bayes (NB) is 0.3 and 0.99, respectively. When compared to prevailing techniques, the anticipated approach also delivers promising outcomes.

Unsupervised Deep Learning Approach for In-Vehicle Intrusion Detection System

Abstract: The controller area network (CAN) is a standard communication protocol used for sending messages between electronic control unit of a modern automotive system. CAN protocol does not have any in-built security mechanisms and, hence, various attacks can affect the vehicle and cause life threats to the passengers. This article presents an unsupervised deep learning architecture for detecting intrusions on a CAN bus. The CAN intrusion detection system (IDS) architecture has an autoencoder that helps to learn the optimal features from CAN packets to differentiate between the normal and attacks. The optimal features are passed as input to the Gaussian mixture model, which helps us to cluster the CAN network packet data samples into normal and attacks. A detailed analysis of the proposed architecture is done on the CAN IDS dataset. To develop a robust CAN IDS system and achieve generalization, the proposed method is evaluated on the other two computer network intrusion datasets and a wireless sensor network dataset. In all the experiments, the proposed method has performed better than the existing unsupervised method and mainly showed a performance gain of 6.4% on the CAN IDS dataset. This indicates that the proposed method is robust and generalizable across detecting various attacks in a CAN bus and most importantly, the method can be used in real time to effectively monitor the CAN network traffic to proactively alert possible attacks.

A hybrid machine learning approach for prediction of conversion from mild cognitive impairment to dementia

Abstract: Mild cognitive impairment (MCI) represents a precursor to dementia for many individuals; however, some forms of MCI tend to remain stable over time and do not progress to dementia. In fact, conversion rates vary substantially depending on the diagnostic criteria used and the nature of the analytic sample and clinical setting. To identify personalized strategies to prevent or slow the progression of dementia and to support the clinical development of novel treatments, we need to develop new approaches for modelling disease progression that can differentiate between progressive and non-progressive MCI subjects. The aim of this study was to develop a novel prognostic machine learning (ML) framework utilising longitudinal information encoded in efficient, cost-effective, and non-invasive markers to identify MCI subjects that are at risk for developing dementia. Our approach was developed using the dataset from the National Alzheimer’s Coordinating Center. We built two prognostic models based on the patient data from 3 (n = 768) (Model 1) and 4 (n = 409) (Model 2) assessment visits. A novel hybrid prognostic approach, using cognitive trajectory classes, generated through unsupervised learning (Stage 1), as input in supervised ML models (Stage 2), was developed and systematically tested. Our unsupervised learning approach (Stage 1) involved: (i) the implementation of the longitudinal data partitioning method allowing for clustering trajectories based on their shapes; (ii) validation of the optimal number of clusters using three different Clustering Validity Indices (CVIs), and (iii) application of the fusion-based methods for combining CVIs into the fused normalized CVI scores, averaged for each cluster partition to determine the final number of trajectory classes for each type of clinical scores. In Stage 2, we built four types of prognostic models based on random forest (RF), Support Vector Machines (SVM), logistic regression (LR), and kNN ensemble approaches. Classification models incorporating both clinical scores and cognitive trajectory classes input showed up to 6.5 % higher accuracy than models based only on clinical scores (p < 0.05 in all cases). Given the patient data from three time points (Model 1), the highest recorded prediction accuracy was achieved for the ensemble and RF model, i.e., 85.0 % (standard deviation: 3.1 %) and 84.6 % (4.1 %) respectively. Using the patient data from four time points (Model 2), the highest accuracy was reported for RF and ensemble models, i.e., 87.5 % (6.1 %) and 86.8 % (3.7 %) respectively. We showed that the incorporation of the output of unsupervised learning significantly improved the performance of supervised ML models. Our prognostic framework can be applied to improve recruitment in clinical trials and to select early interventions for individuals at high risk of developing dementia.

Self-Supervised Remote Sensing Feature Learning: Learning Paradigms, Challenges, and Future Works

Abstract: Deep learning has achieved great success in learning features from massive remote sensing images (RSIs). To better understand the connection between three feature learning paradigms, which are unsupervised feature learning (USFL), supervised feature learning (SFL), and self-supervised feature learning (SSFL), this paper analyzes and compares them from the perspective of feature learning signals, and gives a unified feature learning framework. Under this unified framework, we analyze the advantages of SSFL over the other two learning paradigms in RSI understanding tasks and give a comprehensive review of existing SSFL works in RS, including the pre-training dataset, self-supervised feature learning signals, and the evaluation methods. We further analyze the effects of SSFL signals and pre-training data on the learned features to provide insights into RSI feature learning. Finally, we briefly discuss some open problems and possible research directions.

PAUSE: principled feature attribution for unsupervised gene expression analysis

Abstract: As interest in using unsupervised deep learning models to analyze gene expression data has grown, an increasing number of methods have been developed to make these models more interpretable. These methods can be separated into two groups: post hoc analyses of black box models through feature attribution methods and approaches to build inherently interpretable models through biologically-constrained architectures. We argue that these approaches are not mutually exclusive, but can in fact be usefully combined. We propose PAUSE ( https://github.com/suinleelab/PAUSE ), an unsupervised pathway attribution method that identifies major sources of transcriptomic variation when combined with biologically-constrained neural network models.

Mel Spectrogram-based advanced deep temporal clustering model with unsupervised data for fault diagnosis

Abstract: Fault diagnosis of mechanical equipment using data-driven machine learning methods has been developed recently as a promising technique for improving the reliability of industrial systems. However, these methods suffer from data sparsity due to the difficulty in data collection, which limits the feature extraction of anomalies. To solve this problem, we propose the mel spectrogram-based advanced deep temporal clustering (ADTC) model, which can extract and verify the features of unlabeled data through an unsupervised learning based autoencoder and the K-means. In addition, the ADTC model uses the proposed centroid based learning to obtain calibrated unsupervised learning data by minimizing the data point and target centroid distances for misclustered encoder output features in ensemble-based unsupervised learning. The classifier of the ADTC model uses a supervised learning based deep support vector machine network model, which is robust to nonlinear data, to diagnose the faults of the mechanical equipment. The proposed ADTC model was validated using mechanical equipment dataset with data augmentation to address the imbalanced dataset problem. During experiments, the mel spectrogram-based ADTC model exhibited the best performance in the various industrial environment with a prediction accuracy as high as 98.06%, outperforming other compared algorithms.

Unsupervised Learning for Feature Selection: A Proposed Solution for Botnet Detection in 5G Networks

Abstract: The world has seen exponential growth in deploying Internet of Things (IoT) devices. In recent years, connected IoT devices have surpassed the number of connected non-IoT devices. The number of IoT devices continues to grow and they are becoming a critical component of the national infrastructure. IoT devices' characteristics and inherent limitations make them attractive targets for hackers and cyber criminals. Botnet attack is one of the serious threats on the Internet today. This article proposes pattern-based feature selection methods as part of a machine learning (ML)-based botnet detection system. Specifically, two methods are proposed: the first is based on the most dominant pattern feature values and the second is based on maximal frequent itemset mining. The proposed feature selection method uses Gini impurity and an unsupervised clustering method to select the most influential features automatically. The evaluation results show that the proposed methods have improved the performance of the detection system. The developed system has a true positive rate of 100% and a false positive rate of 0% for best performing models. In addition, the proposed methods reduce the computational cost of the system as evidenced by the detection speed of the system.

Self-Weighted Unsupervised LDA

Abstract: As a hot topic in unsupervised learning, clustering methods have been greatly developed. However, the model becomes more and more complex, and the number of parameters becomes more and more with the continuous development of clustering methods. And parameter-tuning in most methods is a laborious work due to its complexity and unpredictability. How to propose a concise and beautiful model in which the parameters can be learned adaptively becomes a very meaningful problem. Aim at tackling this problem, we develop a novel self-weighted unsupervised linear discriminative analysis method, namely SWULDA. The proposed method not only avoids adjusting parameters but also explains the link between k -means and linear discriminant analysis (LDA). To obtain superior structural performance, the idea of minimizing the within-class scatter matrix and maximizing the between-class scatter matrix is embedded in the unsupervised model. Moreover, equipped with the proposed quadratic weighted optimization framework, the parameter can be adaptively learned. The extensive experiments on several datasets are conducted to validate the effectiveness of our method.

Semi-Supervised and Unsupervised Deep Visual Learning: A Survey

Abstract: State-of-the-art deep learning models are often trained with a large amount of costly labeled training data. However, requiring exhaustive manual annotations may degrade the model's generalizability in the limited-label regime.Semi-supervised learning and unsupervised learning offer promising paradigms to learn from an abundance of unlabeled visual data. Recent progress in these paradigms has indicated the strong benefits of leveraging unlabeled data to improve model generalization and provide better model initialization. In this survey, we review the recent advanced deep learning algorithms on semi-supervised learning (SSL) and unsupervised learning (UL) for visual recognition from a unified perspective. To offer a holistic understanding of the state-of-the-art in these areas, we propose a unified taxonomy. We categorize existing representative SSL and UL with comprehensive and insightful analysis to highlight their design rationales in different learning scenarios and applications in different computer vision tasks. Lastly, we discuss the emerging trends and open challenges in SSL and UL to shed light on future critical research directions.

Unsupervised Learning of Interacting Topological Phases from Experimental Observables

Abstract: Classifying topological phases of matter with strong interactions is a notoriously challenging task and has attracted considerable attention in recent years. In this paper, we propose an unsupervised machine learning approach that can classify a wide range of symmetry-protected interacting topological phases directly from the experimental observables and without a priori knowledge. We analytically show that Green’s functions, which can be derived from spectral functions that can be measured directly in an experiment, are suitable for serving as the input data for our learning proposal based on the diffusion map. As a concrete example, we consider a one-dimensional interacting topological insulators model and show that, through extensive numerical simulations, our diffusion map approach works as desired. In addition, we put forward a generic scheme to measure the spectral functions in ultracold atomic systems through momentum-resolved Raman spectroscopy. Our work circumvents the costly diagonalization of the system Hamiltonian, and provides a versatile protocol for the straightforward and autonomous identification of interacting topological phases from experimental observables in an unsupervised manner.

Learning Salient Feature for Salient Object Detection Without Labels

Abstract: Supervised salient object detection (SOD) methods achieve state-of-the-art performance by relying on human-annotated saliency maps, while unsupervised methods attempt to achieve SOD by not using any annotations. In unsupervised SOD, how to obtain saliency in a completely unsupervised manner is a huge challenge. Existing unsupervised methods usually gain saliency by introducing other handcrafted feature-based saliency methods. In general, the location information of salient objects is included in the feature maps. If the features belonging to salient objects are called salient features and the features that do not belong to salient objects, such as background, are called nonsalient features, by dividing the feature maps into salient features and nonsalient features in an unsupervised way, then the object at the location of the salient feature is the salient object. Based on the above motivation, a novel method called learning salient feature (LSF) is proposed, which achieves unsupervised SOD by LSF from the data itself. This method takes enhancing salient feature and suppressing nonsalient features as the objective. Furthermore, a salient object localization method is proposed to roughly locate objects where the salient feature is located, so as to obtain the salient activation map. Usually, the object in the salient activation map is incomplete and contains a lot of noise. To address this issue, a saliency map update strategy is introduced to gradually remove noise and strengthen boundaries. The visualization of images and their salient activation maps show that our method can effectively learn salient visual objects. Experiments show that we achieve superior unsupervised performance on a series of datasets.

High-throughput cryo-ET structural pattern mining by unsupervised deep iterative subtomogram clustering.

Abstract: Cryoelectron tomography directly visualizes heterogeneous macromolecular structures in their native and complex cellular environments. However, existing computer-assisted structure sorting approaches are low throughput or inherently limited due to their dependency on available templates and manual labels. Here, we introduce a high-throughput template-and-label-free deep learning approach, Deep Iterative Subtomogram Clustering Approach (DISCA), that automatically detects subsets of homogeneous structures by learning and modeling 3D structural features and their distributions. Evaluation on five experimental cryo-ET datasets shows that an unsupervised deep learning based method can detect diverse structures with a wide range of molecular sizes. This unsupervised detection paves the way for systematic unbiased recognition of macromolecular complexes in situ.

Unsupervised Learning of Disentangled Representation via Auto-Encoding: A Survey

Abstract: In recent years, the rapid development of deep learning approaches has paved the way to explore the underlying factors that explain the data. In particular, several methods have been proposed to learn to identify and disentangle these underlying explanatory factors in order to improve the learning process and model generalization. However, extracting this representation with little or no supervision remains a key challenge in machine learning. In this paper, we provide a theoretical outlook on recent advances in the field of unsupervised representation learning with a focus on auto-encoding-based approaches and on the most well-known supervised disentanglement metrics. We cover the current state-of-the-art methods for learning disentangled representation in an unsupervised manner while pointing out the connection between each method and its added value on disentanglement. Further, we discuss how to quantify disentanglement and present an in-depth analysis of associated metrics. We conclude by carrying out a comparative evaluation of these metrics according to three criteria, (i) modularity, (ii) compactness and (iii) informativeness. Finally, we show that only the Mutual Information Gap score (MIG) meets all three criteria.

Unsupervised ECG Analysis: A Review

Abstract: Electrocardiography is the gold standard technique for detecting abnormal heart conditions. Automatic detection of electrocardiogram (ECG) abnormalities helps clinicians analyze the large amount of data produced daily by cardiac monitors. As thenumber of abnormal ECG samples with cardiologist-supplied labels required to train supervised machine learning models is limited, there is a growing need for unsupervised learning methods for ECG analysis. Unsupervised learning aims to partition ECG samples into distinct abnormality classes without cardiologist-supplied labels–a process referred to as ECG clustering. In addition to abnormality detection, ECG clustering has recently discovered inter and intra-individual patterns that reveal valuable information about the whole body and mind, such as emotions, mental disorders, and metabolic levels. ECG clustering can also resolve specific challenges facing supervised learning systems, such as the imbalanced data problem, and can enhance biometric systems. While several reviews exist on supervised ECG systems, a comprehensive review of unsupervised ECG analysis techniques is still lacking. This study reviews ECG clustering techniques developed mainly in the last decade. The focus will be on recent machine learning and deep learning algorithms and their practical applications. We critically review and compare these techniques, discuss their applications and limitations, and provide future research directions. This review provides further insights into ECG clustering and presents the necessary information required to adopt the appropriate algorithm for a specific application.

On continuous health monitoring of bridges under serious environmental variability by an innovative multi-task unsupervised learning method

Abstract: Design of an automated and continuous framework is of paramount importance to structural health monitoring (SHM). This study proposes an innovative multi-task unsupervised learning method for early assessment of damage in large-scale bridge structures under long-term monitoring. This method entails three main tasks of data cleaning, data partitioning, and anomaly detection. The first task includes discarding missing data and providing outlier-free samples by developing an approach based on the well-known DBSCAN algorithm. Accordingly, this approach enforces the DBSCAN to generate two clusters, one of which contains outlier-free samples and the other one comprises outlier data. In the second task, the outlier-free samples are fed into spectral clustering to partition them into local clusters. Subsequently, a cluster with the maximum cumulative local density is selected as the optimal partition whose features are extracted as the representative data. Finally, local empirical measures under the theory of empirical learning are used to compute anomaly indices for SHM. Long-term modal frequencies of two full-scale bridges are incorporated to verify the proposed method alongside comparative analyses. Results prove that this method can effectively detect damage by providing discriminative anomaly scores and mitigating the negative influences of severe environmental variability.

Machine Learning Assisted Clustering of Nanoparticle Structures

Abstract: We propose a scheme for the automatic separation (i.e., clustering) of data sets composed of several nanoparticle (NP) structures by means of Machine Learning techniques. These data sets originate from atomistic simulations, such as global optimizations searches and molecular dynamics simulations, which can produce large outputs that are often difficult to inspect by hand. By combining a description of NPs based on their local atomic environment with unsupervised learning algorithms, such as K-Means and Gaussian mixture model, we are able to distinguish between different structural motifs (e.g., icosahedra, decahedra, polyicosahedra, fcc fragments, twins, and so on). We show that this method is able to improve over the results obtained previously thanks to the successful implementation of a more detailed description of NPs, especially for systems showing a large variety of structures, including disordered ones.

MDFlow: Unsupervised Optical Flow Learning by Reliable Mutual Knowledge Distillation

Abstract: Recent works have shown that optical flow can be learned by deep networks from unlabelled image pairs based on brightness constancy assumption and smoothness prior. Current approaches additionally impose an augmentation regularization term for continual self-supervision, which has been proved to be effective on difficult matching regions. However, this method also amplify the inevitable mismatch in unsupervised setting, blocking the learning process towards optimal solution. To break the dilemma, we propose a novel mutual distillation framework to transfer reliable knowledge back and forth between the teacher and student networks for alternate improvement. Concretely, taking estimation of off-the-shelf unsupervised approach as pseudo labels, our insight locates at defining a confidence selection mechanism to extract relative good matches, and then add diverse data augmentation for distilling adequate and reliable knowledge from teacher to student. Thanks to the decouple nature of our method, we can choose a stronger student architecture for sufficient learning. Finally, better student prediction is adopted to transfer knowledge back to the efficient teacher without additional costs in real deployment. Rather than formulating it as a supervised task, we find that introducing an extra unsupervised term for multi-target learning achieves best final results. Extensive experiments show that our approach, termed MDFlow, achieves state-of-the-art real-time accuracy and generalization ability on challenging benchmarks. Code is available at https://github.com/ltkong218/MDFlow.