Search

Predicting stock market trends is the basic daily routine task that investors should perform in the stock trading market. Traditional market trends prediction models are generally based on hand-crafted factors or features, which heavily rely on expensive expertise knowledge. Moreover, it is difficult to discover hidden features contained in the stock time series data, which are otherwise helpful for predicting stock market trends. In this paper, we propose a novel stock market trends prediction framework SMART with a self-supervised stock technical data sequence embedding model S3E. Specifically, the model encodes stock technical data sequences into embeddings, which are further trained with multiple self-supervised auxiliary tasks. With the learned sequence embeddings, we make stock market trends predictions based on an LSTM and a feed-forward neural network. We conduct extensive experiments on China A-Shares market and NASDAQ market to show that our model is highly effective for stock market trends prediction. We further deploy SMART in a leading financial service provider in China and the result demonstrates the effectiveness of the proposed method in real-world applications. [ABSTRACT FROM AUTHOR]

Simion Zoo is a Reinforcement Learning (RL) workbench developed for training of novel users that can be deployed over computer farms allowing extensive experimentation over distributed resources. In this paper, we present this software platform and share some of the insights gained during the development. This workbench provides a complete set of tools to design, run, carry out statistical analysis of the results, report preparation, and have qualitative visual assessment of the simulation evolution of continuous RL control experiments. The main features that set apart Simion Zoo from other software packages for introduction to RL experimentation are its easy-to-use GUI, its support for distributed execution including deployment over graphics processing units (GPUs), and the possibility to explore concurrently the RL hyper-parameter space, which is key to successful RL experimentation. [ABSTRACT FROM AUTHOR]

In this paper, a novel discrete-time interconnected model composed of nodes and edges subsystems is proposed, to depict the complex dynamical networks (CDNs) consisted of nodes and dynamic edges which are coupled mutually. Firstly, we propose a vector model to describe the dynamic characteristic of edges rather than the matrix form, the proposed discrete-time CDNs have dynamics attached on not only the nodes but also the edges. Then, to reveal the essential mechanism between network structure and stabilization (synchronization), the control strategies and the coupling modes between two subsystems are designed such that the stabilization and synchronization can be achieved. Finally, a simulation example is given to prove the effectiveness of the proposed methods. [ABSTRACT FROM AUTHOR]

In this paper, a novel decentralized control method is proposed for nonlinear mismatched large-scale interconnected systems subjected to partially unknown dynamics by designing auxiliary control for each subsystem. It is demonstrated that the control sequence consisting of the optimal control policies of auxiliary control can stabilize the system asymptotically, leading to decentralized control of the large-scale system. An integral reinforcement learning (IRL) method is firstly proposed, replacing the traditional policy iterative algorithm to analyze the optimal control problem of each edge subsystem with partially unknown dynamics. After that, the edge-based dynamic event-triggered control algorithm is proposed based on the static event-triggered control method, and an internal dynamic variable characterized by a first-order filter is defined. A single critic neural network (NN) is then designed to learn the approximate optimal control strategy under the dynamic event-triggered mechanism adaptively. The stability analysis is proposed to demonstrate that the state of the event-based pulse system is ultimately uniformly bounded (UUB) and the Zeno behavior is eliminated successfully. Finally, the effectiveness of the proposed algorithm is verified by two simulation examples to realize decentralized control of mismatched large-scale systems. • Edge dynamic event-triggered control is first time combined with IRL algorithms. • The IRL algorithm relaxes the requirement of the system drift dynamics. • Edge dynamic event-triggered method has better performance than static triggering. • Solving the decentralized optimal control problem of mismatched large-scale systems. • The dead zone operations are utilized in this work. [ABSTRACT FROM AUTHOR]

In general, infinity-norm optimization (INO) is commonly deemed as a subset of nonlinear optimization. In the last decade, there have been relatively few reports on solving INO problems, specifically from the time-dependent aspect originating from the real-time motion planning of robots. Therefore, to compensate for the vacancy, this paper proposes a class of neural dynamics (ND) solvers utilizing an ACP framework that combines artificial systems (A), computational experiments (C), and parallel execution (P), respectively. Specifically, two improved ND solvers are constructed by exploiting simplified sign-bi-power and saturation activation functions for solving time-dependent INO (TDINO) problems subject to equality and inequality constraints. Moreover, the corresponding theoretical analysis and proof are conducted, which ensures that residual errors of the proposed improved ND solvers based on the ACP framework converge in a short time. Compared with a SOTA (state-of-the-art) zeroing neural network model presented recently, the average error of the proposed ones is reduced by 52%, the speed of error convergence is increased by 152%, and the transient state is extended by 4% on average. Finally, simulation results of an illustrative example and an application in the robot are provided, which illustrates the effectiveness, superiority, and feasibility of the proposed ND solvers. [ABSTRACT FROM AUTHOR]

In view of zero-sum differential games, this paper addresses the fault-tolerant tracking control (FTTC) problem for nonlinear systems with multiplicative actuator faults. By augmenting the nominal system, the trajectory tracking problem is transformed into an optimal regulation problem. A fault observer is designed to estimate the multiplicative actuator fault factor. In order to better reflect system performance, a comprehensive performance index function including the trajectory tracking error, the control input, and the multiplicative actuator fault is constructed. By considering the control input and the multiplicative actuator fault as two players in the zero-sum differential game (ZSDG), the FTTC problem is thus transformed into a ZSDG problem. Then, a fault-tolerant control based on ZSDG is designed by solving the Hamilton–Jacobi-Isaacs (HJI) equation. Since the HJI equation is difficult to solve, a critic neural network is constructed to obtain its approximation, which constitutes the Nash equilibrium of the ZSDG. The closed-loop faulty system is guaranteed to be uniformly ultimately bounded via the Lyapunov's direct method. The effectiveness of the developed FTTC method is demonstrated by two simulation examples. [ABSTRACT FROM AUTHOR]

Vehicle re-identification (re-ID) has received increasing attention due to its tremendous potential in practical intelligent transportation scenarios. The main difficulties of vehicle re-ID are generally induced by insufficient annotations, varying viewpoints and illuminations, and mismatched visual-semantic cues. Previous studies on vehicle re-ID have generally focused on exploring better visual representations of the target samples while ignoring semantic complementarity from external sources. In this paper, we propose an unsupervised semantic transfer method based on collaborative matrix representation (STCMR) for vehicle re-ID that obtains rich semantic representations by learning knowledge from external annotated data. STCMR contains three components: an asymmetrical collaborative matrix representation term, a low-rank regularized relaxed attribute term, and a dual-graph Laplacian regularization term. In particular, we first build two reliable feature matrices to better transfer external annotated semantics from the source domain to the target re-ID domain. We then propose a robust collaborative matrix representation term to decompose the feature matrices into user-defined, domain-shared, and domain-unique parts in an asymmetrical manner. To tackle samples with incomplete semantics, we design a low-rank regularized attribute relaxed term that fully exploits existing annotations to complete missing entries in a soft constraint manner. Finally, we exploit the local geometric structure consistency in source and target domains by a dual-graph Laplacian regularization term to adaptively estimate the relationships of source domain samples. In addition, we further propose an alternating iterative algorithm to solve the optimization problem. Experimental results conducted on the VehicleID and VeRi-776 datasets validate the effectiveness of our proposed method. [ABSTRACT FROM AUTHOR]

Recurrent neural networks (RNNs) are a popular family of models widely used when facing sequential data such as videos. However, RNNs make assumptions about state transitions that could be damageable. This paper presents two theoretical limitations of RNNs along with popular extensions proposed to mitigate those issues. The effectiveness of these methods is assessed in practice on the specific task of sign language (SL) video tokenization, as it remains challenging. Evaluated strategies enhance transition modeling with RNNs functioning as state machines. However, this performance gain diminishes in more complex architectures, indicating there is still room for improvement. Such improvement would help to build powerful SL tokenizers usable in future pipelines in natural language processing. • Sign language tokenization is a challenging task involving complex transitions. • The limitations of RNNs highlighted in theory can be observed in practice. • Extensions have a positive impact, but there is still room for improvement. • Sign language tokenization task could benefit from these improvements. • There is an interest in a better transition modeling in recurrent neural networks. [ABSTRACT FROM AUTHOR]

With the widespread use of distributed machine learning (DML), many IT companies have established networks dedicated to DML. Different communication architectures of DML have different traffic patterns and different requirements on network performance, which is closely related to network topology. However, traditional network topologies usually pursue general goals and are agnostic to the special communication pattern of the applications. The mismatch between network topology and the applications will directly affect the training performance. Although some studies have analyzed the effect of topology on training performance, the topologies and communication architectures involved are not comprehensive, and it is still not known which topology is appropriate for which communication architecture. This survey investigates typical topologies and analyzes whether they meet the requirements of three commonly used communication architectures (i.e., Parameter Server (PS), Tree and Ring architectures) of DML. Specifically, the topology requirements of each communication architecture and two common topology requirements (i.e., high scalability and fault tolerance) for DML are studied firstly. Next, whether these topologies meet the topology requirements is analyzed. Then, this paper discusses potential technologies and approaches to construct the appropriate scheme for each topology requirement, and then presents DMLNet , a novel network topology that suits the three communication architectures. Finally, several potential directions for future research are outlined. [ABSTRACT FROM AUTHOR]

In human–computer dialogue systems, intent recognition is crucial for determining the intentions or purposes of users during interactions with the system, enabling the system to provide appropriate responses or actions. This paper proposes an intent recognition model that integrates sequential information and sentence structural features. Specifically, the approach utilizes a CNN to capture local salient features in the text, followed by a BILSTM to extract sequential information within the local context. The sequential information is then fed into a multi-head attention mechanism to focus on more relevant sequential details. Additionally, the original data is processed by BERT to extract sentence structural features. Finally, the sequential information features and sentence structural features are concatenated and fused to achieve enhanced intent recognition performance. This approach effectively leverages contextual and semantic information within the text, leading to improved accuracy in intent recognition. Experimental results demonstrate the effectiveness of the proposed method in intent recognition and its high relevance for practical applications. [ABSTRACT FROM AUTHOR]

Data augmentation is a crucial and challenging task for improving defect detection with limited data. Many generative models have been proposed and shown promising performance on this task. However, existing models are unable to capture the fine features of defects when training data is scarce, resulting in the inability to synthesize defects and a lack of diversity in the synthesized defects. Additionally, most models do not consider the location of synthesized defects in the image, thus limiting the ability for augmenting defect data through data generation. In this paper, we propose a new augmentation model named Scarce Data Augmentation Generative Adversarial Nets (Scarcity-GAN) to address the scarce data augmentation problem. Firstly, we design a new clustering module which selects data containing similar features to the target defect from extra datasets, in order to help the GAN learn the features of the target defect. Secondly, we modify the vanilla generator with an Encoder–Decoder model. The generator takes two inputs: one is the defect-free images, which are encoded by the Encoder to obtain defect-free features, and the other is the extra defect feature maps in the target defect set after clustering. Next, we design a Fusion Patch-Embedding module to merge the two different features, ensuring that the synthesized defects are located on the object accurately. We also design a new loss function for the generator, and then prove that it makes our model get converged. Last, we conduct extensive experiments to demonstrate the significant performance improvement and generalizability of Scarcity-GAN on two scarce datasets: industrial O-ring and Metal Iron Sheet datasets; and one general dataset: the public CODEBRIM dataset. The experimental results show that our Scarcity-GAN outperforms the SOTA augmentation models on different scarce datasets. [ABSTRACT FROM AUTHOR]

Knowledge distillation (KD) is a method that transfers information from a larger network (teacher) to a smaller network (student) to obtain stronger performance without extra computational load. It has made great success in image classification, but only attains trivial improvement in object detection. In this paper, we propose a knowledge distillation scheme for object detection with Inconsistency-Based Feature Imitation (IBFI) and Global Relation Imitation (GRI) schemes. IBFI calculates the difference between the outputs of classification head and regression head to balance the classification and localization abilities of the detector. GRI enables the student to mimic the teacher's relational information. Extensive experiments have been conducted on popular datasets including MS COCO and PASCAL VOC to validate the effectiveness of our scheme. We achieve 39.9% mAP on MS COCO by using RetinaNet + ResNet-50, which surpasses the baseline by 2.5%. [ABSTRACT FROM AUTHOR]

In this paper, the finite-time quasi-synchronization of a multi-layer heterogeneous network with a leader node is studied. A novel distributed control approach that incorporates impulsive control is introduced to achieve finite-time quasi-synchronization of the whole network. The control protocol is distributed, the impulses are synchronized, and the maximum time interval of impulses is bounded. Within this framework and based on the maximum impulsive interval, a finite-time quasi-synchronization criterion for the multi-layer heterogeneous network is established. A close relationship between the settling time and the maximum impulsive interval is discovered. Simulations are shown for illustration and verification. [ABSTRACT FROM AUTHOR]

Inference centers need more data to have a more comprehensive and beneficial learning model, and for this purpose, they need to collect data from data providers. On the other hand, data providers are cautious about delivering their datasets to inference centers in terms of privacy considerations. In this paper, by modifying the structure of the autoencoder, we present a method that manages the utility-privacy trade-off well. To be more precise, the data is first compressed using the encoder, then confidential and non-confidential features are separated and uncorrelated using the classifier. The confidential feature is appropriately combined with noise, and the non-confidential feature is enhanced, and at the end, data with the original data format is produced by the decoder. The suggested architecture additionally enables data providers to modify the degree of privacy needed for private features and the level of utility for non-private features. The proposed method has been examined for both image and categorical databases, and the results show a significant performance improvement compared to previous methods. [ABSTRACT FROM AUTHOR]

Tremendous progress has been made in facial Micro-Expression (ME) spotting and recognition; however, most works have focused on either spotting or recognition tasks on the 2D videos. Until recently, the estimation of the 3D motion field (a.k.a scene flow) for the ME has only become possible after the release of the multi-modal ME dataset. In this paper, we propose the first Scene Flow Attention-based Micro-expression Network, namely SFAMNet. It takes the scene flow computed using the RGB-D flow algorithm as the input and predicts the spotting confidence score and emotion labels. Specifically, SFAMNet is an attention-based end-to-end multi-stream multi-task network devised to spot and recognize the ME. Besides that, we present a data augmentation strategy to alleviate the small sample size problem during network learning. Extensive experiments are performed on three tasks: (i) ME spotting; (ii) ME recognition; and (iii) ME analysis on the multi-modal CAS(ME) 3 dataset. Empirical results indicate that depth is vital in capturing the ME information and the effectiveness of the proposed approach. Our source code is publicly available at https://github.com/genbing99/SFAMNet. • First study to spot and recognize micro-expressions on the multi-modal dataset. • A scene flow attention-based end-to-end multi-stream multi-task network is proposed. • Scene flow is extracted to estimate the 3D motion using color and depth modalities. • This work leverages data augmentation and improves the pseudo-labeling technique. • Encouraging results in micro-expression spotting, recognition, and analysis. [ABSTRACT FROM AUTHOR]

Bio-inspired neuromorphic cameras asynchronously record pixel brightness changes and generate sparse event streams. They can capture dynamic scenes with little motion blur and more details in extreme illumination conditions. Due to the multidimensional address-event structure, most existing vision algorithms cannot properly handle asynchronous event streams. While several event representations and processing methods have been developed to address such an issue, they are typically driven by a large number of events, leading to substantial overheads in runtime and memory. In this paper, we propose a new graph representation of the event data and couple it with a Graph Transformer to perform accurate neuromorphic classification. Extensive experiments show that our approach leads to better results and excels at the challenging realistic situations where only a small number of events and limited computational resources are available, paving the way for neuromorphic applications embedded into mobile facilities. • Asynchronous event streams of neuromorphic cameras shape synchronous graph representations. • Graph Transformer effectively learns informative features from event-based graphs. • Temporal-wise connectivity and learning suffer less from the loss of information. • Event-based graphs feature lower runtime and smaller memory footprints. [ABSTRACT FROM AUTHOR]

Supervised learning-based code smell detection has become one of the dominant approaches to identify code smell. Existing works optimize the process of code smell detection from multiple aspects, such as high-quality dataset, feature selection, and model, etc. Although the accuracy is improved continuously, researchers are confused about what model are the most suitable ones to detect code smell when considering dataset construction and feature selection. Furthermore, existing surveys for code smell mainly analyze the impact of code smell, categorize the concerns of code smell, and repair code smell. There is a lack of systematic analysis and classification of code smell detection based on supervised learning. To this end, we collect 86 papers of code smell detection based on supervised learning ranging from January 2010 to April 2023. A total of 7 research questions is empirically evaluated from different aspects, such as datasets construction, data pre-processing, feature selection, and model training, etc. We conclude that existing works suffer from issues such as sample imbalance, different attention to types of code smell, and limited feature selection. Finally, we suggest possible future research directions. [ABSTRACT FROM AUTHOR]

When recovering the sharp image from a blurry observation, moving objects, e.g. , people and vehicles, usually attract more perceptual attention. However, existing motion deblurring methods primarily focus on removing the global camera motion blur, while neglecting the object motion blur. In this paper, we propose a j oint l earning framework for m otion deblurring and d efocus deblurring (JLMD), where a motion deblurring network can be trained for removing object motion blur, and a defocus deblurring network can be also obtained as a byproduct. Especially, we propose to circumvent the hardship in obtaining all-sharp ground truth for object motion blurry images by leveraging the complementary nature between object motion blur and defocus blur. The JLMD framework incorporates three key components. Firstly, a mask generation module is deployed to predict the boundary of motion blur. Secondly, a mutual supervision mechanism is elaborated, which enables the defocus blurry image and the motion blurry image to serve as the ground truth for each other. Thirdly, a content preservation loss is designed to encourage non-blurry regions to stay unchanged. Furthermore, we introduce a Defocus and Motion Deblurring Dataset (DMDD), which is the first real-world dataset specifically dedicated to object motion blurry images. Extensive experiments validate that our method is more effective in removing object motion blur than state-of-the-art methods, while it is at least comparable or superior in removing global camera motion blur and defocus blur. The source code and dataset are available at https://github.com/liyucs/JLMD. [ABSTRACT FROM AUTHOR]

Neural networks are widely used in various applications such as image classification, speech recognition and natural language processing. As intelligent systems often rely on neural networks to make critical decisions, it is crucial to ensure that the behavior of a neural network is trustworthy. Along with the increase in the scale and complexity of neural networks, it becomes a challenging problem to verify their safety, since the large-scale state space may make the verification of safety properties infeasible or unsolvable. In order to achieve efficient safety verification of neural networks, this paper proposes a divide-and-conquer verification method based on neuron importance analysis. The method converts the safety constraints to be verified into an optimization problem composed of output neurons, and recursively measures the influence of neurons in the network on the final objective function to calculate their importance. The neurons with the greatest importance are split to refine the constraints, so that the original verification problem is transformed into a set of sub-problems for solution. The method is implemented as a tool named NIAVerify, which is evaluated through experiments of property verification on ReLU-based fully-connected benchmark networks trained on two classic datasets. Experimental results show that NIAVerify can achieve significant state space reduction and the verification efficiency is 44.80% higher than the existing verification method. [ABSTRACT FROM AUTHOR]

Multimodal sentiment analysis (MSA) aims to detect the sentiments from language (text), audio, and visual (facial expressions) modalities. The main challenge in MSA is how to efficiently model intra-modality and inter-modality dynamics. With the advent of graph convolution network (GCN), graph-based models are proposed to solve the challenge. However, general graphs contain only two nodes per edge, which limits the exploitation of high-order interactions. Moreover, current graph-based models mainly aggregate the features of each node during fusion, while the features of connected edges are not well mined. In this paper, we introduce dynamic hypergraph convolution networks to MSA for the first time and propose a Multimodal Dynamic Hypergraph Network (MDH) to learn intra- and inter-modality dynamics. Hypergraphs provide a natural approach to capture transcendental pairwise relations, and their potential for MSA remains unexplored. MDH mainly consists of three components: Unimodal Encoder, Dynamic Hypergraph Enhancement Network (DHEN), and HyperFusion module. Specifically, DHEN is composed of Cross-modal Affine, Hypergraph Construction, and Hypergraph Aggregation modules. As for the intra-modality dynamics, MDH utilizes Hypergraph Construction and Aggregation modules to model the interactions within time steps for each modality. As for the inter-modality dynamics, MDH implements Cross-modal Affine and HyperFusion modules to learn the relationships of the modalities. In addition, multi-task learning has been implemented to optimize the learning process for multimodal tasks. Experiments show that MDH outperforms graph-based models on CMU-MOSI and CMU-MOSEI datasets, as well as obtains new state-of-the-art results on CH-SIMS dataset. Furthermore, we conduct external experiments to explore the effectiveness of MDH and the effect of model depth with different graph networks. [ABSTRACT FROM AUTHOR]

Feature selection is a vital preprocessing step in real applications of data mining and machine learning. With the prevalence of high-dimensional hybrid data sets in real-world scenarios, along with the presence of test costs and misclassification costs, the need for effective feature selection methods has become more prominent. However, existing feature selection approaches mainly focus on cost-sensitive data from a single granularity perspective, and they are primarily applicable to single-typed data sets. To address these limitations, this paper presents a novel feature selection approach specifically designed for hybrid data, considering variable test costs and misclassification costs. The proposed method is based on neighborhood multigranulation rough sets (NMRS), which provides more comprehensive and multi-angle data analysis for cost-sensitive hybrid data. First, a novel multigranulation model is developed to effectively process cost-sensitive hybrid data. Building upon this model, a cost-based multi-criteria measure is proposed to evaluate the significance of features. This measure takes into account the comprehensive information of candidate features, including their power in algebraic view, information view, and associated costs. Furthermore, a heuristic feature selection algorithm based on NMRS is proposed to handle hybrid data with the test costs and misclassification costs. This algorithm leverages the benefits of the proposed multigranulation model and cost-based measure to identify the most discriminative features efficiently. Finally, the experimental results on twelve different datasets show that the proposed heuristic algorithm outperforms other compared algorithms, especially in total cost and classification accuracy. [ABSTRACT FROM AUTHOR]

Numerous disjoint community detection methods have reached the state-of-the-art. Some overlapping community detection methods have been proposed in recent years, but they lack the ability to adjust the degree of overlap while maintaining detection quality. To well handle this issue, we in this paper propose a novel method, namely expansion with contraction method for overlapping community detection (ECOCD). Specifically, ECOCD obtains the disjoint communities through non-negative matrix factorization and proceeds to expansion with contraction process (including the expansion process and the contraction process). In each iteration of the process, we randomly select a community and then continuously conduct the expansion and contraction processes on this community. The former process absorbs nodes by the degree of affiliation that is newly defined, while the latter removes nodes by permanence. Moreover, we theoretically analyze the computational complexity of ECOCD. The advantage of ECOCD is that it is applicable to various networks with different properties by adjusting the degree of overlap, and enjoys high quality of overlapping community detection as well. Our experiments on both synthetic and real-world networks further verify this. Extensive experiments show that ECOCD is superior to the eleven state-of-the-art overlapping community detection methods in terms of four metrics, validating the effectiveness, efficiency and robustness of ECOCD. [ABSTRACT FROM AUTHOR]

Aspect category sentiment analysis (ACSA) excels at identifying the aspect categories and corresponding sentiments involved in a sentence, regardless of whether the aspect terms are explicitly mentioned or not. However, current methods tend to overinflate the original data, resulting in the introduction of unnecessary information, and fail to capture the inter-task relationship sufficiently. This paper presents a new method termed the prompt-based joint model (PBJM) to address these complications. PBJM treats the sentiment polarity prediction as binary classification and leverages a natural language prompt template, a concise sentence that guides the model to perform aspect category identification subtask and curtails the need for data augmentation. The two subtasks are jointly trained in pre-trained language models (PLMs) to capture their correlation. Further, the attention mechanism for aspect categories enables the model to concentrate selectively on significant features such as phrases and words during the predictions. In addition, the verbalizer employs a set of parameters to balance the weight of each label word while projecting between the label space and the label words space. Through experiments on four datasets, our model demonstrated remarkable performance in detecting category-sentiment pairs. [ABSTRACT FROM AUTHOR]

As the deep neural networks (DNNs) with satisfied expression ability usually require large scale to gain adequate performance, and deploying large-scale DNNs on resource-limited environments is still a challenge, neural network compression becomes a hot topic nowadays. Among the multiple compression methods, tensor decomposition reveals many specific advantages, such as regular data structure comes from linear algebra, convenient approach of training from scratch, ideal compression ratio, etc. Nevertheless, for some compact neural modules such as two-dimensional/three-dimensional (2D/3D) convolutional kernels, traditional tensor decomposition in the way of approximation has encountered intractable obstacles. Fortunately, some works utilize the tensor mapping approach, which just regards the data structure of tensor decomposition as neural layers, to reconstruct the convolutional kernel into a new lightweight module with several thinner convolutional kernels. The only two flies in the ointment are there lack necessary theories of tensor mapping, and there is still no tensor mapping way to compress high-order three-dimensional convolutional neural networks (3DCNNs). In this paper, we first deeply analyze the tensor mapping theory including convergence and precision, which separately establishes the rationality of tensor mapping and its superiority over the traditional tensor approximation, according to the Lottery Ticket Hypothesis. Then we propose an efficient method termed as 3D-KCPNet, to compress 3DCNNs based on the novel Kronecker canonical polyadic (KCP) tensor decomposition. The proposed method can not only compress the 3D convolutional kernels, but also convert a 3D convolution to efficient 1 × 1 × 1 and 2D depthwise convolutions. The experiments on the video recognition datasets VIVA Challenge, UCF11, UCF50, and UCF101 show that the accuracy of 3D-KCPNet can surpass its original baseline model and the corresponding tensor approximation model. [ABSTRACT FROM AUTHOR]

Reinforcement learning is capable of providing state-of-art performance in end-to-end robotic control tasks. Nevertheless, many real-world control tasks necessitate the balancing of multiple conflicting objectives while simultaneously ensuring that the learned policies adhere to constraints. Additionally, individual users may typically prefer to explore the personalized and diversified robotic control modes via specific preferences. Therefore, this paper presents a novel constrained multi-objective reinforcement learning algorithm for personalized end-to-end robotic control with continuous actions, allowing a trained single model to approximate the Pareto optimal policies for any user-specified preferences. The proposed approach is formulated as a constrained multi-objective Markov decision process, incorporating a nonlinear constraint design to facilitate the agent in learning optimal policies that align with specified user preferences across the entire preference space. Meanwhile, a comprehensive index based on hypervolume and entropy is presented to measure the convergence, diversity and evenness of the learned control policies. The proposed scheme is evaluated on nine multi-objective end-to-end robotic control tasks with continuous action space, and its effectiveness is demonstrated in comparison with the competitive baselines, including classical and state-of-the-art algorithms. [Display omitted] • A constrained multi-objective reinforcement learning approach is proposed. • The method is formulated as a constrained multi-objective Markov decision process. • An evenness metric is designed through Shannon–Wiener diversity index. [ABSTRACT FROM AUTHOR]

Unmanned-Aerial-Vehicles (UAVs) represent an active research topic over multiple fields for performing inspection, delivery and surveillance applications among other operations. However, achieving the utmost efficiency requires drones to perform these tasks without the need of human intervention, which demands a robust and accurate localization system for achieving a safe and efficient autonomous navigation. Nevertheless, currently used satellite-based localization systems like GPS are insufficient for high-precision applications, especially in harsh scenarios like indoor and deep urban environments. In these contexts, Local Positioning Systems (LPS) have been widely proposed for satisfying the localization requirements of these vehicles. However, the performance of LPS is highly dependent on the actual localization architecture and the spatial disposition of the deployed sensor distribution. Therefore, before the deployment of an extensive localization network, an analysis regarding localization architecture and sensor distribution should be taken into consideration for the task at hand. Nonetheless, no actual study is proposed either for comparing localization architectures or for attaining a solution for the Node Location Problem (NLP), a problem of NP-Hard complexity. Therefore, in this paper, we propose a comparison among synchronous LPS for determining the most suited system for localizing UAVs over urban scenarios. We employ the Cràmer–Rao-Bound (CRB) for evaluating the performance of each localization system, based on the provided error characterization of each synchronous architecture. Furthermore, in order to attain the optimal sensor distribution for each architecture, a Black-Widow-Optimization (BWO) algorithm is devised for the NLP and the application at hand. The results obtained denote the effectiveness of the devised technique and recommend the implementation of Time Difference Of Arrival (TDOA) over Time of Arrival (TOA) systems, attaining up to 47% less localization uncertainty due to the unnecessary synchronization of the target clock with the architecture sensors in the TDOA architecture. • The comparison among different time-based localization architectures for UAVs navigation in deep urban environments. • A novel analysis for optimally deploying time-based localization systems for drone localization in urban scenarios. • The proposal of a BWO algorithm for attaining both the NLP and the SSP problem with different localization architectures. [ABSTRACT FROM AUTHOR]

In this paper, a dynamic compensator-based near-optimal control approach for unknown nonaffine nonlinear systems is developed by using integral reinforcement learning. Since system dynamics is unknown, it is difficult to obtain the optimal control policy via neuro-dynamic programming. To address this problem, a general dynamic compensator is introduced as the virtual control input to augment the unknown nonaffine nonlinear system as a partially unknown affine system. For the augmented system, a novel quadratic value function is designed with the system states, the actual control input and the virtual control input. The optimal control of the augmented system can be regarded as the near-optimal control for the original system since the novel optimal value function is an upper bound of the original optimal value function. In order to avoid the identification of system dynamics, the integral reinforcement learning framework is utilized to derive the optimal control based on the solution of Hamilton–Jacobi–Bellman equation via the critic-only structure. Meanwhile, the weight learning rule of the critic neural network is presented with the experience replay technique to relax the persistence of excitation condition. Moreover, the uniform ultimate boundedness of weight estimation errors and the stability of the closed-loop system are guaranteed by using the Lyapunov's direct method. Finally, simulation results of two examples demonstrate the effectiveness of the developed dynamic compensator-based near-optimal control method. [ABSTRACT FROM AUTHOR]

Spiking neural network is used to solve the problems of power consumption and computation requirement of the traditional artificial neural network. This paper proposes an event-driven multilayer spiking convolutional neural network (SCNN) hardware system for the image classification task. The proposed hardware system consists of a microcontroller unit (MCU) and an SCNN processor. Input images are temporally encoded to spike trains in the form of address event representation (AER) by MCU, and sent to the SCNN processor to execute AER-based spike-centric convolution in order to reduce redundant operations and get the classification result efficiently. By rearranging the storage structure of weights of synapses and membrane potential (MP) of spiking neurons, multi-channel parallel MP updating is realized to reduce latency. In addition, FIFOs are inserted between the spiking layers to construct a pipeline structure. The hardware architecture of the proposed system is implemented on the Xilinx ZYNQ-7000 platform. At the operating frequency of 100 MHz, the proposed SCNN processor has a speed of 400classifications/s and 80uJ/inference. [ABSTRACT FROM AUTHOR]

Hydraulic systems present numerous challenges in the development of high-performance tracking controllers due to their highly nonlinear characteristics and heavy modeling uncertainties. Addressing this issue, this paper presents a friction compensation-based finite-time tracking control strategy with prescribed performance constraints. The system model is first derived, taking into consideration the nonlinear behaviors and potential uncertainties encountered in practical scenarios. Then, an adaptive finite-time prescribed performance controller is synthesized based on the backstepping framework. A novel form of finite-time performance function is introduced to constrain the tracking error, which provides the closed-loop system with a less complex guarantee of finite-time convergence and can also deliver the required transient performance and steady-state accuracy. Additionally, given the impact of uncertainty terms and the difficulty in obtaining the virtual command derivative, we employ neural networks (NNs) to estimate unknown dynamics and introduce command filters to acquire the intermediate signals, simplifying the backstepping control design process. Theoretical analysis indicates that the proposed control strategy can achieve the desired tracking performance and ensure the stability of the entire closed-loop system. Comparative numerical simulations are presented to confirm the usefulness of the suggested approach. • Efficient Implementation of Finite-time Prescribed Performance Neural Network Control Scheme. • Effective Friction Compensation Strategy for Servo Systems. • Design and Application of Low-Complexity Controllers. [ABSTRACT FROM AUTHOR]

• Token-level and term-level interactions to handle various overlapped triplets. • Latent graph parsing structure iteratively refines vital token-level interactions. • Term pairing captures complex interactions between terms effectively. Aspect sentiment triplet extraction task detects three elements of fine-grained sentiment analysis from given sentence, including aspect and opinion terms and their sentiment polarity. Existing methods mainly include tagging-based and span-based methods, where the former show defects on handling overlapped triplets, and the latter could theoretically handle all overlapped triplets but lack of tailored inter-word dependency so that suffer from insufficient span semantic. In this paper, we propose a span-based dependency-enhanced graph convolutional network, which leverages contextual semantic and latent dependency to enrich span representations. Specifically, we devise a latent graph convolutional network to emphasize critical inter-word dependencies and cut off redundant connections in a learnable gating manner, improving the information flow during inter-word interaction. In addition, considering the problem of multi-word term sentiment consistency, we detect effective aspect and opinion terms derived from the output of span enumeration, and introduce term-level interactions by coupling, which meanwhile enables our model to deal with various types of triplets including many-to-one and one-to-many overlapped triplets. Extensive experiments over four benchmark datasets verify that the proposed method outperforms all the baselines with an average F1 improvement of up to 6.13%, and meanwhile shows fine interpretability. The experimental results demonstrate that effectively enhancing token-level and term-level interactions can significantly improve the aspect sentiment triplet extraction performance. [ABSTRACT FROM AUTHOR]

This paper proposes optimal bipartite consensus issue for discrete-time multi-agent system (DTMAS) using event-triggered control (ETC) and quantized event-triggered control (QETC). Firstly, a novel event-triggered control strategy is introduced along with specific conditions to ensure the stability for DTMAS. Due to the intractability of the Hamilton–Jacobi–Bellman (HJB) equation, through two neural networks (NNs), the value function and optimal control are approximated by the adaptive dynamic programming (ADP) method. The weight matrix of the action neural network is updated only when triggered, while also taking into account the quantization effect of the information. The stability of the weights errors and the convergence of the system dynamic are demonstrated using Lyapunov stability theory. Finally, the practicality of the improved method is confirmed a numerical example. [ABSTRACT FROM AUTHOR]

This paper is devoted to proposing a novel dynamic event-triggered adaptive dynamic programming (ADP) algorithm to tackle the tracking control problem of unknown nonlinear systems with uncertain input saturation. Initially, an augmented system is established to address the investigated optimal tracking control issue with a discounted cost function. An online identifier based on neural networks (NNs) is designed to recover the unknown system dynamics. Furthermore, the controller design consists of two parts, including the optimal control policy for the nominal system and the NN-based feedforward compensator. In particular, the latter leverages adaptive NN techniques to cope with uncertain input constraints which can be deemed as uncertain saturation nonlinearities. A critic NN is constructed to obtain the approximate optimal control policy for the nominal system. Moreover, rather than traditional time-triggered and static event-triggered control schemes, an adaptive dynamic event-triggering condition is designed to determine the system sampling as well as the controller execution to further reduce transmission and computation burden. The introduction of an internal dynamic variable in possession of a positive constant lower bound can realize the dynamic adjustment of triggering threshold and effectively exclude Zeno behavior. Notice that the weights of identifier NN and critic NN are updated simultaneously, only at the event-triggering instants. Subsequently, the Lyapunov-based theoretical analysis is conducted to confirm the uniform ultimate boundedness of the tracking error and all the NN weight estimation errors. Eventually, the feasibility and effectiveness of the developed algorithm are validated by means of two simulation examples. [ABSTRACT FROM AUTHOR]

The challenge to automatically design machine learning-based models for datasets with varying dimensions (features) remains open, especially in the context of unstructured and noisy data. Bayesian neural networks employ Markov chain Monte Carlo (MCMC) and variational inference methods for training (sampling) model parameters. However, the progress of MCMC methods in deep learning has been slow due to high computational requirements and uninformative priors of model parameters. Reversible jump MCMC allows sampling of model parameters of variable lengths; hence, it has the potential to train Bayesian neural networks effectively. In this paper, we implement reversible jump MCMC sampling for training dynamic Bayesian neural networks that feature cascaded neural networks with dynamic hidden and input neurons. Our dynamic Bayesian neural network can model unstructured data of varying dimensions and provide uncertainty quantification in model predictions. We apply the methodology to selected regression and classification problems from the literature. The results show that our method provides an effective approach for the dynamic exploration of models while featuring uncertainty quantification that not only caters to model parameters but also extends to model topology. This opens up the road for uncertainty quantification in dynamic deep learning models for dynamic and unstructured data streams. [ABSTRACT FROM AUTHOR]

Distributed fiber optic vibration sensing system has been widely used in safety monitoring with distinct advantages, and the feature extraction and classification methods of fiber optic signals directly determine the real-time performance and reliability of the monitoring system. The existing research feature extraction methods are unidimensional and time consuming, which cannot balance the goals of high accuracy and low time consumption for safety monitoring systems. In this paper, we propose an efficient recognition framework for fusing signal time–frequency features, called TFF-CNN, based on Gramian Angular Difference Fields(GADF) and FFT co-generation matrix(FFTT) to manually extract two-dimensional time–frequency image features of fiber optic signals, taking the significant advantages of CNN in image processing and combining a two-channel model and a fusion module that simulates human decision-making behavior. The accuracy of TFF-CNN is 99.30 % and the detection response time is only 0.6 s. Compared with other methods in this field, TFF-CNN has the advantages of low false alarm rate and short time consumption, which is more suitable for deployment in security monitoring field with distributed fiber optic sensing system. Index Terms : Distributed optical fiber vibration sensing system(DVS), Two-dimensional multi-feature, CNN, intrusion detection, real-time monitoring. [ABSTRACT FROM AUTHOR]

In this paper, a quaternion-variable zeroing neurodynamic (ZND) is designed, in order to address a class of quaternion-variable time-varying convex optimization problems (TCOP) with equality constraints and general inequality constraints. Through utilizing the Karush–Kuhn–Tucker (KKT) condition, the TCOP is transformed into a time-varying equation system for solving. During this process, the introduction of two approaches distinguishes our works in solving TCOPs. On the one hand, the penalty approach is utilized to reduce the influence of time-varying inequality constraints, and thus, makes the ZND more suitable for TCOPs with general and complex inequality constraints. On the other hand, the approach to simplify the KKT condition significantly simplifies the structure of the proposed ZND by reducing the neurons' number. The proposed ZND realizes efficient convergence while keeping the structure as simple as possible, making it easy to apply in engineering tasks. Besides, to improve its efficiency further, a nonlinear activated function is utilized to achieve a faster convergent rate, by which the convergent rate of the ZND is accelerated to finite-time convergence. Apart from that, the ZND here also owes superior robustness. To the best of our knowledge, this is the first time to extend the ZND for solving TCOP to the quaternion domain. Finally, numerical examples and a robot manipulator control experiment further explain the effectiveness of the proposed ZND. [ABSTRACT FROM AUTHOR]

The point cloud semantic segmentation plays a significant role in the understanding of 3D environment. However, current 3D point cloud segmentation methods pay little attention in the inevitable blurring and loss of point information within deep network, which considerably impairs the segmentation performance, particularly in the complex and colorful indoor scenes. To overcome the inevitable loss of point information (position, color and normal vector), in this paper, we propose a brand-new Information Encoding and Fusion Module (IEFM), comprising a point Information Encoding Method (IEM) and an optimized Multi-Encoding Fusion Method (EFM). In terms of constructing the innovative bias standardization of point information and effectively merging other information encodings into position encoding, IEFM adaptively complements the loss of the point information and thereby achieving enhanced environment perception ability. Additionally, although the proposed IEFM is capable of managing multi-class point information, the encoding interference resulting from the coexistence of multiple information is quite unavoidable, leading to detrimental consequences for the overall semantic segmentation performance. Therefore, to reduce the information encoding interference problem, we further innovatively propose the Information Distribution Strategy (IDS), in terms of hierarchically distribute kinds of point information, so that the interference of multiple information encodings will be extremely mitigated and achieving more accurate indoor point cloud semantic segmentation. Benefiting from the modular design, the proposed IEFM and IDS can be easily inserted in existing point-based point cloud segmentation models. The experimental results have shown the effectiveness of our proposed methods across multiple state-of-the-art models and benchmarks (S3DIS and ScanNet), achieving competitive performance of 77.4% mIoU on the large-scale S3DIS benchmark. • The loss of 3D point cloud information in deep network is inevitable. • Introducing information encoding can effectively complement the lost information. • Flexibly distribute information can alleviate encoding interference problem. [ABSTRACT FROM AUTHOR]

The increasing complexity and amount of data available in modern applications strongly demand Trustworthy Learning algorithms that can be fed directly with complex and large graphs data. In fact, on one hand, machine learning models must meet high technical standards (e.g., high accuracy with limited computational requirements), but, at the same time, they must be sure not to discriminate against subgroups of the population (e.g., based on gender or ethnicity). Graph Neural Networks (GNNs) are currently the most effective solution to meet the technical requirements, even if it has been demonstrated that they inherit and amplify the biases contained in the data as a reflection of societal inequities. In fact, when dealing with graph data, these biases can be hidden not only in the node attributes but also in the connections between entities. Several Fair GNNs have been proposed in the literature, with uNIfying Fairness and stabiliTY (NIFTY) (Agarwal et al., 2021) being one of the most effective. In this paper, we will empower NIFTY's fairness with two new strategies. The first one is a Biased Edge Dropout, namely, we drop graph edges to balance homophilous and heterophilous sensitive connections, mitigating the bias induced by subgroup node cardinality. The second one is Attributes Preprocessing, which is the process of learning a fair transformation of the original node attributes. The effectiveness of our proposal will be tested on a series of datasets with increasingly challenging scenarios. These scenarios will deal with different levels of knowledge about the entire graph, i.e., how many portions of the graph are known and which sub-portion is labelled at the training and forward phases. • We considered NIFTY, one of the most effective Fair Graph Neural Networks • We empowered it with two new strategies • Biased Edge Dropout balances homophilous and heterophilous sensitive connections • Fair Attribute Preprocessing learns a fair transformation of the node attributes • We tested the proposal on several datasets and scenarios [ABSTRACT FROM AUTHOR]

In this paper, systems of the fractional-order multidimension-valued memristive neural networks (FMVMNNs) with leakage delay and time-varying delays are studied in continuous-time and discrete-time cases. Comparative studies on synchronizations control between discrete-time FMVMNNs and the continuous-time system are investigated directly rather than through the decomposition method. By employing Lyapunov theory and fractional derivative inequalities, sufficient criteria about the quasi-synchronization and global ultimate Mittag-Leffler lag quasi-synchronization are obtained in multidimension-valued linear matrix inequality form, while succinct conclusions of asymptotical synchronization and global Mittag-Leffler synchronization are derived in vector form. The obtained criteria have many advantages in a less computation, lower conservatism, more generality and higher flexibility. Finally, four examples are given to demonstrate the effectiveness and availability of the derived conclusions. [ABSTRACT FROM AUTHOR]

State-of-the-art SLAM methods are designed to work only with the type of camera employed to create the map, and little attention has been paid to the reusability of the maps created. In other words, the maps generated by current methods can only be reused with the same camera employed to create them. This paper presents a novel SLAM approach that allows maps generated with one camera to be used by other cameras with different resolutions and optics. Our system allows, for instance, creating highly detailed maps processed off-line with high-end computers, to be reused later by low-powered devices (e.g. a drone or robot) using a different camera. The first map, called base map, can be reused with other cameras and dynamically adapted by creating an augmented map. The principal idea of our method is a bottom-up pyramidal representation of the images that allows us to match keypoints between different camera types seamlessly. The experiments conducted validate our proposal, showing that it outperforms the state-of-the-art approaches, namely ORBSLAM, OpenVSLAM and UcoSLAM. • Proposed new SLAM system able to work with multiple and heterogeneous cameras. • An internal bottom-up image representation allows comparing keypoints of cameras with different optics and resolutions. • Different cameras can reuse a base map created with a camera. • Comparison with state-of-the-art methods shows better performance. [ABSTRACT FROM AUTHOR]

• Context learning extraction from ship trajectory data. • AIS data real-world data use for data mining problems and anomaly detection. • Trajectory compression and segmentation techniques. • Data mining techniques for trajectory clustering applications. This paper presents a context information extraction process over Automatic Identification System (AIS) real-world ship data, building a system with the capability to extract representative points of a trajectory cluster. With the trajectory cluster, the study proposes the use of trajectory segmentation algorithms to extract representative points of each trajectory and then use the k-means algorithm to obtain a series of centroids over all the representative points. These centroids, combined, form a new representative trajectory of the cluster. This new representative trajectory of the input cluster represents new contextual information extracted from the original set of trajectories, being possible to apply anomaly detection approaches over the new obtained context. The results show a suitable approach with several compression algorithms that are compared with a metric based on the Perpendicular Euclidean Distance. [ABSTRACT FROM AUTHOR]

Transformer-based methods have currently demonstrated impressive results in the field of skeleton-based action recognition. Nevertheless, how to effectively model multi-scale features with transformers remains a challenging problem, which is crucial to distinguish various actions. In this paper, we propose a Space–time Dual Multi-scale transformer (STDM-transformer) to explore the multi-scale collaborative representation employing both fine and coarse scale motion information. In contrast to existing approaches which typically propagate information between scales in a single fusion manner, our Space–time Dual Multi-scale method stratifies the space–time multi-scale into dual levels. One level is to construct fine-grained local motion interactions. In detail, the space–time multi-scale partition strategy and the novel intra-inter space–time transformer module are proposed to extract and aggregate the feature in part scale and body scale, respectively. The other is aimed at modeling coarse-grained global motion contexts, in which the layer-wise multi-scale progressive fusion strategy is designed. Extensive experimental results demonstrate that the proposed STDM-transformer achieves the SOTA performance on large-scale datasets. • We propose a space-time dual multi-scale transformer for action recognition. • The Space-Time partition strategy and the intra-inter transformer block are proposed. • The layer-wise multi-scale progressive fusion strategy is designed. • Our method achieves SOTA performance on large scale datasets. [ABSTRACT FROM AUTHOR]

The growing potential of multimodal short videos has contributed to a new type of recommendation. It depends on effectively measuring the similarities between the short video pairs, which consist of video frames and descriptions. Previous studies proposed using contrastive learning with in-batch negative sampling. However, such a strategy would take other semantically similar samples in the batch as negatives, leading to a biased issue. This paper proposes a debiased momentum contrastive learning (DMCL) on the unified multimodal Transformer model (VideoSim) for video similarity measures. The proposed DMCL alleviates the bias issue of negative sampling by introducing implicit knowledge of the model itself as soft labels. Instead of simply taking other samples in the batch as negatives, the proposed DMCL applies soft labels as supervision from the ground truth and the contextual semantic similarities between video-text pairs to alleviate the bias caused by negative sampling. In addition, DMCL expands the negative samples by a momentum queue strategy, which allows storing more multimodal representations to contrast more negatives. The experimental results of measuring multimodal video similarity show that the proposed DMCL outperforms all baselines in terms of Spearman's rank correlation. Ablation studies and extensive analysis further demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

The weapon-target assignment problem is a challenging optimization issue, but reliability is seldom considered in the majority of existing literature. To address the high-reliability weapon-target assignment problem, this paper integrates weapon reliability and mission reliability into a multi-objective optimization model (MOD) and presents an improved algorithm termed MOEA/D-iAM2M to the problem. This algorithm effectively combines the strengths of adaptive search space decomposition-based MOEA (MOEA/D-AM2M) and two-stage hybrid learning-based MOEA (HLMEA), resulting in a faster convergence rate and a more extensive distribution of the Pareto solution set. Furthermore, a reference point is incorporated into MOEA/D-iAM2M to facilitate the adaptive weight adjustment. Numerical experiments are carried out to confirm the effectiveness of the proposed MOEA/D-iAM2M. This research is significant in the field of optimization and has practical value in the defense industry. [ABSTRACT FROM AUTHOR]

Deep Supervised Salient Object Detection (SSOD) excessively relies on large-scale annotated pixel-level labels which consume intensive labour acquiring high quality labels. In such precondition, deep Unsupervised Salient Object Detection (USOD) draws public attention. Under the framework of the existing deep USOD methods, they mostly generate pseudo labels by fusing several hand-crafted detectors' results. On top of that, a Fully Convolutional Network (FCN) will be trained to detect salient regions separately. While the existing USOD methods have achieved some progress, there are still challenges for them towards satisfactory performance on the complex scene, including (1) poor object wholeness owing to neglecting the hierarchy of those salient regions; (2) unsatisfactory pseudo labels causing by unprimitive fusion of hand-crafted results. To address these issues, in this paper, we introduce the property of part-whole relations endowed by a Belief Capsule Network (BCNet) for deep USOD, which is achieved by a multi-stream capsule routing strategy with a belief score for each stream within the CapsNets architecture. To train BCNet well, we generate high-quality pseudo labels from multiple hand-crafted detectors by developing a consistency-aware fusion strategy. Concretely, a weeding out criterion is first defined to filter out unreliable training samples based on the inter-method consistency among four hand-crafted saliency maps. In the following, a dynamic fusion mechanism is designed to generate high-quality pseudo labels from the remaining samples for BCNet training. Experiments on five public datasets illustrate the superiority of the proposed method. Codes have been released on: https://github.com/Mirlongue/Deep-Unsupervised-Part-Whole-Relational-Visual-Saliency. [ABSTRACT FROM AUTHOR]

This paper presents a human–robot interaction system (HRIS) that utilizes human perception and action recognition to enable the robot to understand human intentions and flexibly interact with humans. A monocular multi-person three-dimensional (3D) pose estimation method is first proposed to perceive multi-person two-dimensional (2D) and 3D poses in interaction scenarios. Furthermore, a 3D skeleton poses tracking approach is adopted to locate the identity of each person in consecutive frames and enhance interactive stability. Then, an action recognition model is developed, which exploits tracked pose features to recognize the intentions of humans. An action-controlled interaction system is built with a modular approach to ensure flexibility in meeting multiple task requirements and facilitating flexible interaction. In the system, a distance-based safety solution is designed to avoid collisions between humans and robots. Finally, experimental results are presented to demonstrate the feasibility and effectiveness of the proposed methods and system. [ABSTRACT FROM AUTHOR]

The bias in machine learning models has gained increasing attention in recent years, as these models can reflect and even amplify biases present in the data used to train them. One approach to mitigating bias is identifying and down-weight features that contribute disproportionately to model predictions, which can be accomplished using saliency techniques. Current debiasing methods often lead to the loss of contextual information, where the model tends to respond incorrectly even when the gender information is present in the context; hence, even though the bias reduces, performance (coreference resolution, fluency) also reduces. This paper explores data augmentation and saliency techniques to mitigate bias in natural language generation. Specifically, we explore applying the saliency technique of SHAP (SHapley Additive exPlanations) over a model trained on debiasing using data augmentation (switching gendered words with counterparts) and then applying hard debiasing to remove the influential biased token. We build a dialogue context test setup to evaluate bias and context relevance using the presence of gendered words in the model-generated responses. The response is evaluated based on the gender information from context to ensure the model follows the gender in context. We demonstrate that this approach can effectively reduce the impact of biased features on model predictions while preserving overall model accuracy. Additionally, we discuss potential limitations and future directions for research in this area. Our findings suggest that saliency offers an avenue to address machine learning bias. [ABSTRACT FROM AUTHOR]