Context-based offline meta-reinforcement learning (meta-RL) is a paradigm that integrates meta-learning with offline reinforcement learning. It learns a strategy to extract task-specific contexts from trajectories of meta-training tasks and leverages this strategy for adapting to unseen target tasks. However, existing methods struggle to generate generalizable contexts for adaptations due to context shift, which arises from the context-based policy overfitting to offline data. We argue that leveraging the internal relationships among tasks, rather than treating each task in isolation, is crucial for mitigating the impact of context shift. Hence, we propose a framework called cross-task contexts for improving generalization in meta-RL (CTMRL). Specifically, we design a context quantization variational auto-encoder (CQ-VAE), which clusters task-specific contexts of meta-training tasks into discrete codes based on the internal relationships among tasks. Cross-task contexts are constructed with these codes, reflecting shared information across similar tasks. These cross-task contexts not only serve as high-level structures to capture similarity across tasks but also provide a foundation for hard contrastive learning that enhances the distinguishability of similar yet distinct tasks, thereby improving the generalization of contexts and facilitating adaptation to unseen target tasks. The evaluation in meta-environments confirms the performance advantage of CTMRL over existing methods.

Meta-learning for Bayesian optimization accelerates optimization by leveraging knowledge from previous tasks, but existing methods optimize for average performance and fail on challenging outlier tasks critical in practice. These limitations become particularly severe when target tasks exhibit distribution shifts or when optimization budgets are limited in real-world applications. We introduce MetaGameBO, a hierarchical game-theoretic framework that formulates meta-learning as robust optimization through CVaR-based task selection and diversity-aware sample learning. Our approach incorporates uncertainty-aware adaptation via probabilistic embeddings and Thompson sampling for robust generalization to out-of-distribution targets. We establish theoretical guarantees including convergence to game-theoretic equilibria and improved sample complexity, and demonstrate substantial improvements with 95.7% reduction in average loss and 88.6% lower tail risk compared to state-of-the-art methods on challenging tasks and distribution shifts.

Neural processes (NPs) are a promising paradigm to enable skill transfer learning across tasks with the aid of the distribution of functions. The previous NPs employ the empirical risk minimization principle in optimization. However, the fast adaption ability to different tasks can vary widely, and the worst fast adaptation can be catastrophic in risk-sensitive tasks. To achieve robust neural processes modeling, we consider the problem of training models in a risk-averse manner, which can control the worst fast adaption cases at a certain probabilistic level. By transferring the risk minimization problem to a two-level finite sum minimax optimization problem, we can easily solve it via a double-looped stochastic mirror prox algorithm with a task-aware variance reduction mechanism via sampling samples across all tasks. The mirror prox technique ensures better handling of complex constraint sets and non-Euclidean geometries, making the optimization adaptable to various tasks. The final solution, by aggregating prox points with the adaptive learning rates, enables a stable and high-quality output. The proposed learning strategy can work with various NPs flexibly and achieves less biased approximation with a theoretical guarantee. To illustrate the superiority of the proposed model, we perform experiments on both synthetic and real-world data, and the results demonstrate that our approach not only helps to achieve more accurate performance but also improves model robustness.

The growing demand for efficient deep learning has positioned dataset distillation as a pivotal technique for compressing training dataset while preserving model performance. However, existing inner-loop optimization methods for dataset distillation typically rely on random truncation strategies, which lack flexibility and often yield suboptimal results. In this work, we observe that neural networks exhibit distinct learning dynamics across different training stages—early, middle, and late—making random truncation ineffective. To address this limitation, we propose Automatic Truncated Backpropagation Through Time (AT-BPTT), a novel framework that dynamically adapts both truncation positions and window sizes according to intrinsic gradient behavior. AT-BPTT introduces three key components: (1) a probabilistic mechanism for stage-aware timestep selection, (2) an adaptive window sizing strategy based on gradient variation, and (3) a low-rank Hessian approximation to reduce computational overhead. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that AT-BPTT achieves state-of-the-art performance, improving accuracy by an average of 6.16\% over baseline methods. Moreover, our approach accelerates inner-loop optimization by 3.9 × while saving 63\% memory cost.

This paper focuses on the design of event-triggered controllers for the synchronization of delayed Takagi–Sugeno (T–S) fuzzy neural networks (NNs) under deception attacks. The traditional event-triggered mechanism (ETM) determines the next trigger based on the current sample, resulting in network congestion. Furthermore, such methods suffer from the issues of deception attacks and unmeasurable system states. To enhance the system stability, we adaptively detect the occurrence of events over a period of time. In addition, deception attacks are recharacterized to describe general scenarios. Specifically, the following enhancements are implemented: First, we use a Bernoulli process to model the occurrence of deception attacks, which can describe a variety of attack scenarios as a type of general Markov process. Second, we introduce a sum-based dynamic discrete event-triggered mechanism (SDDETM), which uses a combination of past sampled measurements and internal dynamic variables to determine subsequent triggering events. Finally, we incorporate a dynamic output feedback controller (DOFC) to ensure the system stability. The concurrent design of the DOFC and SDDETM parameters is achieved through the application of the cone complement linearization (CCL) algorithm. We further perform two simulation examples to validate the effectiveness of the algorithm.

Predicting missing facts for temporal knowledge graphs (TKGs) is a fundamental task, called temporal knowledge graph completion (TKGC). One key challenge in this task is the imbalance in data distribution, where facts are unevenly spread across entities and timestamps. This imbalance can lead to poor completion performance for long-tail entities and timestamps, and unstable training due to the introduction of false negative samples. Unfortunately, few previous studies have investigated how to mitigate these effects. Moreover, for the first time, we found that existing methods suffer from model preferences, revealing that entities with specific properties (e.g., recently active) are favored by different models. Such preferences will lead to error accumulation and further exacerbate the effects of imbalanced data distribution. To alleviate the impacts of imbalanced data and model preferences, we introduce Booster , the first data augmentation strategy for TKGs. The unique requirements here lie in generating new samples that fit the complex semantic and temporal patterns within TKGs, and identifying hard-learning samples specific to models. Therefore, we propose a hierarchical scoring algorithm based on triadic closures within TKGs. By incorporating both global semantic patterns and local time-aware structures, the algorithm enables pattern-aware validation for new samples. Meanwhile, we propose a two-stage training approach to identify samples that deviate from the model's preferred patterns. With a frequency-based filtering strategy, this approach also helps to avoid the misleading of false negatives. Experiments justify that Booster can seamlessly adapt to existing TKGC models and achieve on average 4.5% performance improvement.

Transferability and imperceptibility of adversarial examples are pivotal for assessing the efficacy of black-box attacks. While diffusion models have been employed to generate adversarial examples, leveraging their advanced image generation capability to enhance transferability and imperceptibility, these methods typically focus only on perturbing the image or latent space. They often ignore the critical role of semantic information in the denoising process, thereby impeding the improvement of the transferability of adversarial examples. Furthermore, the modification of high-level semantics inevitably introduces image blurring. This degradation in visual quality makes the adversarial examples more susceptible to detection. To overcome the above limitations, we are the first to utilize image latent encoding and semantic embedding perturbations to enhance the performance of adversarial attacks. Then, the LESEP method is proposed. In the LESEP framework, we first apply image latent encoding attack to achieve deception of the target model. Second, the semantic embedding attack enhances the transferability of adversarial examples. Additionally, we utilize the image restoration technique to guarantee the high imperceptibility of the crafted adversarial examples. Through comprehensive experiments on diverse datasets, different network architectures and defense methods, we have demonstrated that the LESEP method achieves outstanding transferability and imperceptibility while displaying strong robustness.

In this paper, we present a novel task of source-free cross-modal adversarial example generation, which generates adversarial examples based on textual descriptions of attackers. This task has two challenges as follows. First, how to generate adversarial examples when the clean examples are missing or inaccessible. Second, how to achieve fine-grained custom adversarial example generation according to the semantic descriptions of the attackers. Existing adversarial example generation methods can not effectively deal with these two challenges. To address these challenges, we propose a Source-Free Cross-Modal Adversarial Example Generation framework, abbreviated as SFCM-AEG. Within the SFCM-AEG model, we firstly leverage a pre-trained GPT as a simulator to construct textual descriptions of attackers by labels. Following this, we employ a diffusion model to synthesize an image that aligns with the generated textual description. Finally, the generated images are converted into adversarial examples using an adversarial example generation method. Experimental results demonstrate that our proposed SFCM-AEG method can generate adversarial examples with customized semantic descriptions, without relying on clean examples, while achieving strong attack performance in a white-box setting.

Training generative adversarial networks (GANs) for noise-to-image synthesis is a challenge task, primarily due to the instability of GANs’ training process. One of the key issues is the generator’s sensitivity to input data, which can cause sudden fluctuations in the generator’s loss value with certain inputs. This sensitivity suggests an inadequate ability to resist disturbances in the generator, causing the discriminator’s loss value to oscillate and negatively impacting the discriminator. Then, the negative feedback of discriminator is also not conducive to updating generator’s parameters, leading to suboptimal image generation quality. In response to this challenge, we present an innovative GANs model equipped with a learnable auxiliary module that processes auxiliary noise. The core objective of this module is to enhance the stability of both the generator and discriminator throughout the training process. To achieve this target, we incorporate a learnable auxiliary penalty and an augmented discriminator, designed to control the generator and reinforce the discriminator’s stability, respectively. We further apply our method to the Hinge and LSGANs loss functions, illustrating its efficacy in reducing the instability of both the generator and the discriminator. The tests we conducted on LSUN, CelebA, Market-1501, and Creative Senz3D datasets serve as proof of our method’s ability to improve the training stability and overall performance of the baseline methods.

Incomplete Multi-View Clustering (IMVC) is a promising topic in multimedia as it breaks the data completeness assumption. Most existing methods solve IMVC from the perspective of graph learning. In contrast, self-representation learning enjoys a superior ability to explore relationships among samples. However, only a few works have explored the potentiality of self-representation learning in IMVC. These self-representation methods infer missing entries from the perspective of whole samples, resulting in redundant information. In addition, designing an effective strategy to retain salient features while eliminating noise is rarely considered in IMVC. To tackle these issues, we propose a novel self-representation learning method with missing sample recovery and enhanced low-rank tensor regularization. Specifically, the missing samples are inferred by leveraging the local structure of each view, which is constructed from available samples at the feature level. Then an enhanced tensor norm, referred to as Logarithm-p norm is devised, which can obtain an accurate cross-view description by adaptive weights. Our proposed method achieves exact subspace representation in IMVC by leveraging high-order correlations and inferring missing information at the feature level. Extensive experiments on several widely used multi-view datasets demonstrate the effectiveness of the proposed method.

The open-vocabulary human-object interaction (Ov-HOI) detection aims to identify both base and novel categories of human-object interactions while only base categories are available during training. Existing Ov-HOI methods commonly leverage knowledge distilled from CLIP to extend their ability to detect previously unseen interaction categories. However, our empirical observations indicate that the inherent noise present in CLIP has a detrimental effect on HOI prediction. Moreover, the absence of novel human-object position distributions often leads to overfitting on the base categories within their learned queries. To address these issues, we propose a two-step framework named, CaM-LQ, Calibrating visual-language Models, (e.g., CLIP) for open-vocabulary HOI detection with Locality-aware Queries. By injecting the fine-grained HOI supervision from the calibrated CLIP into the HOI decoder, our model can achieve the goal of predicting novel interactions. Extensive experimental results demonstrate that our approach performs well in open-vocabulary human-object interaction detection, surpassing state-of-the-art methods across multiple metrics on mainstream datasets and showing superior open-vocabulary HOI detection performance, e.g., with 4.54 points improvement on the HICO-DET dataset over the SoTA CLIP4HOI on the UV task with the same backbone ResNet-50.

Spectrum sharing technology is used to alleviate the tension and scarcity of spectrum resources, and federated learning can significantly enhance the performance of tasks such as incumbent detection and improve the quality of spectrum sharing. However, spectrum sharing methods based on federated learning still face challenges such as large-scale data transmission and the lack of privacy protection for sensing nodes. To tackle these issues, in this paper, we propose a compressed sensing (CS) based transmission framework that integrates efficient aggregation and privacy protection. In particular, a multiple measurement vector (MMV)-CS model is used for efficient aggregation between the central server and sensing nodes. By designing different measurement vectors, local environment sensing nodes can be divided into different clusters, forming multiple superimposed transmission signals at the central server. Thus, the central server will obtain the aggregated information of local models from different clusters, completing the optimization of the global model in federated learning. In this process, the efficiency of data aggregation has been greatly improved, and the data privacy of individual environment sensing nodes is protected. The security analysis and simulation results are provided to validate the effectiveness of the proposed schemes. The detection performance of the proposed method is as good as that of the approach under the raw training samples, while the privacy-preserving and communication efficiency are significantly improved.

Approximating the uncertainty of an emotional AI agent is crucial for improving the reliability of such agents and facilitating human-in-the-loop solutions, especially in critical scenarios. However, none of the existing systems for emotion recognition in conversation (ERC) has attempted to estimate the uncertainty of their predictions. In this article, we present HU-Dialogue, which models hierarchical uncertainty for the ERC task. We perturb contextual attention weight values with source-adaptive noises within each modality, as a regularization scheme to model context-level uncertainty and adapt the Bayesian deep learning method to the capsule-based prediction layer to model modality-level uncertainty. Furthermore, a weight-sharing triplet structure with conditional layer normalization is introduced to detect both invariance and equivariance among modalities for ERC. We provide a detailed empirical analysis for extensive experiments, which shows that our model outperforms previous state-of-the-art methods on three popular multimodal ERC datasets.

Extracting temporal abstraction (option), which empowers the action space, is a crucial challenge in hierarchical reinforcement learning. Under a well-structured action space, decision-making agents can probe more deeply in the searching or plan efficiently through pruning irrelevant action candidates. However, automatically capturing a well-performed temporal abstraction is a nontrivial challenge due to its insufficient exploration and inadequate functionality. We consider alleviating this challenge from two perspectives, i.e., diversity and individuality. For the aspect of diversity, we propose a maximum entropy model based on ensembled options to encourage exploration. For the aspect of individuality, we propose to distinguish each option accurately, utilizing mutual formation minimization, so that each option can better express and function. We name our framework as an ensemble with soft option (ESO) critics. Furthermore, the residual algorithm (RA) with a bidirectional target network is introduced to stabilize bootstrapping, yielding a residual version of ESO. We provide detailed analysis for extensive experiments, which shows that our method boosts performance in commonly used continuous control tasks.

The goal of cross-view geo-localization is to determine the location of a given ground image by matching with aerial images. However, existing methods ignore the variability of scenes, additional information and spatial correspondence of covisibility and non-convisibility areas in ground-aerial image pairs. In this context, we propose a cross-view matching method called SMDT with image alignment and Transformer. First, we utilize semantic segmentation technique to segment different areas. Then, we convert the vertical view of aerial images to front view by mixing polar mapping and perspective mapping. Next, we simultaneously train dual conditional generative adversarial nets by taking the semantic segmentation images and converted images as input to synthesize the aerial image with ground view style. These steps are collectively referred to as image alignment. Last, we use Transformer to explicitly utilize the properties of self-attention. Experiments show that our SMDT method is superior to the existing ground-to-aerial cross-view methods.

The goal of cross-view image matching based on geo-localization is to determine the location of a given ground-view image (front view) by matching it with a group of satellite-view images (vertical view) with geographic tags. Due to the rapid development of unmanned aerial vehicle (UAV) technology in recent years, it has provided a real viewpoint close to 45 degrees (oblique view) to bridge the visual gap between views. However, existing methods ignore the direct geometric space correspondence of UAV-satellite views, and only use brute force for feature matching, leading to inferior performance. In this context, we propose an end-to-end cross-view matching method that integrates cross-view synthesis module and geo-localization module, which fully considers the spatial correspondence of UAV-satellite views and the surrounding area information. To be specific, the cross-view synthesis module includes two parts: the oblique view of UAV is first converted to the vertical view by perspective projection transformation (PPT), which makes the UAV image closer to the satellite image; then we use conditional generative adversarial nets (CGAN) to synthesize the UAV image with vertical view style, which is close to the real satellite image by learning the converted UAV as the input image and the real satellite image as the label. Geo-localization module refers to existing local pattern network (LPN), which explicitly considers the surrounding environment of the target building. These modules are integrated in a single architecture called PCL, which mutually reinforce each other. Our method is superior to the existing UAV-satellite cross-view methods, which improves by about 5%.

Fully mining visual cues to aid in content understanding is crucial for video captioning. However, most state-of-the-art video captioning methods are limited to generating captions purely based on straightforward information while ignoring the scenario and context information. To fill the gap, we propose a novel, simple but effective scenario-aware recurrent transformer (SART) model to execute video captioning. Our model contains a “scenario understanding” module to obtain a global perspective across multiple frames, providing a specific scenario to guarantee a goal-directed description. Moreover, for the sake of achieving narrative continuity in the generated paragraph, a unified recurrent transformer is adopted. To demonstrate the effectiveness of our proposed SART, we have conducted comprehensive experiments on various large-scale video description datasets, including ActivityNet, YouCookII, and VideoStory. Additionally, we extend a story-oriented evaluation framework for assessing the quality of the generated caption more precisely. The superior performance has shown that SART has a strong ability to generate correct, deliberative, and narrative coherent video descriptions.

Multi-sensory data has exhibited a clear advantage in expressing richer and more complex feelings, on the Emotion Recognition in Conversation (ERC) task. Yet, current methods for multimodal dynamics that aggregate modalities or employ additional modality-specific and modality-shared networks are still inadequate in balancing between the sufficiency of multimodal processing and the scalability to incremental multi-sensory data type additions. This incurs a bottleneck of performance improvement of ERC. To this end, we present MetaDrop, a differentiable and end-to-end approach for the ERC task that learns module-wise decisions across modalities and conversation flows simultaneously, which supports adaptive information sharing pattern and dynamic fusion paths. Our framework mitigates the problem of modelling complex multimodal relations while ensuring it enjoys good scalability to the number of modalities. Experiments on two popular multimodal ERC datasets show that MetaDrop achieves new state-of-the-art results.

An observer-based dissipativity control for Takagi-Sugeno (T-S) fuzzy neural networks with distributed time-varying delays is studied in this article. First, the network channel delays are modeled as a distributed delay with its kernel. To make full use of kernels of the distributed delay, a Lyapunov-Krasovskii functional (LKF) is established with the kernel of the distributed delay. It is noted that the novel LKF and delay-dependent reciprocally convex inequality plays an important role in dealing with global asymptotical stability and strict (Q, S,R) - α -dissipativity of the T-S fuzzy delayed model. Through the constructed LKF, a new set of less conservative linear matrix inequality (LMI) conditions is presented to obtain an observer-based controller for the T-S fuzzy delayed model. This proposed observer-based controller ensures that the state of the closed-loop system is globally asymptotically stable and strictly (Q, S,R) - α -dissipative. Finally, the effectiveness of the proposed results is shown in numerical simulations.

The article considers the impulsive synchronization for inertial neural networks with unbounded delay and actuator saturation via sampled-data control. Based on an impulsive differential inequality, the difficulties caused by unbounded delay and impulsive effect may be effectively avoid. By applying polytopic representation technique, the actuator saturation term is first considered into the design of impulsive controller, and less conservative linear matrix inequality (LMI) criteria that guarantee asymptotical synchronization for the considered model via hybrid control are given. As special cases, the asymptotical synchronization of the considered model via sampled-data control and saturating impulsive control are also studied, respectively. Numerical simulations are presented to claim the effectiveness of theoretical analysis. A new image encryption algorithm is proposed to utilize the synchronization theory of hybrid control. The validity of image encryption algorithm can be obtained by experiments.

The impulsive stabilization of nonlinear time-delay system with input saturation via delay-dependent polytopic approach is studied in this article. Different from polytopic representation technique, delay-dependent polytopic technique is able to estimate a larger domain of attraction. Based on this approach, the actuator saturation term is first introduced into the design of impulsive controller, which is expressed as a convex combination of the product of delay-dependent state vectors and auxiliary matrices. By applying delay-dependent polytopic technique and delay-dependent Lyapunov–Krasovskii functional (LKF) approach, a new series of less conservative linear matrix inequality (LMI) criteria are obtained to ensure the stability of the established model. Finally, two examples are presented to claim the effectiveness of theoretical analysis results.

Remote sensing satellite images often suffer from low spatial resolution. Image super-resolution plays an important role in remote sensing image processing. However, existing methods show that increasing network depth will inevitably lead to the dramatic increase of model parameters and the over-fitting problem. Besides, most methods treat different types of information (low-frequency and high-frequency) equally. Motivated by these observations, we propose a fused recurrent network via channel attention (CA-FRN) in this paper. The basic module, recursive channel attention block (RCAB), pays enough attention to the high-frequency information and diminishes the low-frequency information adaptively through channel attention. Based on RCAB, we render our model effective by retaining and fusing hierarchical local information of both low-resolution and high-resolution, and we enhance the network performance simply by increasing the number of RCABs without adding extra parameters. We evaluate the proposed model on satellite images from different datasets, and the proposed CA-FRN is superior to the state-of-the-art methods. Code is available at https://github.com/lxy0922/CAFRN.

Efficient exploration is a major challenge in Reinforcement Learning (RL) and has been studied extensively. However, for a new task existing methods explore either by taking actions that maximize task agnostic objectives (such as information gain) or applying a simple dithering strategy (such as noise injection), which might not be effective enough. In this paper, we investigate whether previous learning experiences can be leveraged to guide exploration of current new task. To this end, we propose a novel Exploration with Structured Noise in Parameter Space (ESNPS) approach. ESNPS utilizes meta-learning and directly uses meta-policy parameters, which contain prior knowledge, as structured noises to perturb the base model for effective exploration in new tasks. Experimental results on four groups of tasks: cheetah velocity, cheetah direction, ant velocity and ant direction demonstrate the superiority of ESNPS against a number of competitive baselines.

Text-based person search aims at retrieving the most relevant pedestrian images from database in response to a query in form of natural language description. Existing algorithms mainly focus on embedding textual and visual features into a common semantic space so that the similarity score of features from different modalities can be computed directly. Softmax loss is widely adopted to classify textual and visual features into a correct category in the joint embedding space. However, softmax loss can only help classify features but not increase the intra-class compactness and inter-class discrepancy. To this end, we propose multiplicative angular margin (MAM) loss to learn angularly discriminative features for each identity. The multiplicative angular margin loss penalizes the angle between feature vector and its corresponding classifier vector to learn more discriminative feature. Moreover, to focus more on informative image-text pair, we propose pairwise similarity weighting (PSW) loss to assign higher weight to informative pairs. Extensive experimental evaluations have been conducted on the CUHK-PEDES dataset over our proposed losses. The results show the superiority of our proposed method. Code is available at https://github.com/pengzhanguestc/MAM_loss.

This article investigates the event-triggered synchronization of delayed neural networks (NNs). A novel integral-based event-triggered scheme (IETS) is proposed where the integral of the system states, and past triggered data over a period of time are used. With the proposed IETS, the integral event-triggered synchronization problem becomes a distributed delay problem. Using the Bessel-Legendre inequalities, sufficient conditions for the existence of a controller that ensures asymptotic synchronization are provided in the form of linear matrix inequalities (LMIs). Illustrative examples are used to demonstrate the advantages of the proposed IETS method over other event-triggered scheme (ETS) methods. Moreover, this IETS method is applied to the image encryption and decryption. A novel encryption algorithm is proposed to enhance the quality of the encryption process.

Person re-identification is a fundamental task in automated video surveillance and has been an area of intensive research in the past few years. Several person re-identification methods based on deep learning have been proposed and achieved remarkable performance. However, extraction of more useful spatial and temporal information from input images and design of a more effective approach to match the same persons are still challenging. In this paper, we present a novel Two-Stage Convolution Neural Network (TSCNN), which effectively extracts the spatio-temporal feature with two-stream network in two directions, and matches the person with a novel convolutional neural network. Extensive experiments are conducted on three public benchmarks, i.e., iLIDS-VID, PRID2011 and MARS datasets. The experimental results demonstrate that the performance of our TSCNN is better in comparison with the state-of-the-art methods. The code of TSCNN is available at https://github.com/zyoohv/TSCNN.

Person re-identification is an important task in video surveillance, focusing on finding the same person across different cameras. However, most existing methods of video-based person re-identification still have some limitations (e.g., the lack of effective deep learning framework, the robustness of the model, and the same treatment for all video frames) which make them unable to achieve better recognition performance. In this paper, we propose a novel self-paced learning algorithm for video-based person re-identification, which could gradually learn from simple to complex samples for a mature and stable model. Self-paced learning is employed to enhance video-based person re-identification based on deep neural network, so that deep neural network and self-paced learning are unified into one frame. Then, based on the trained self-paced learning, we propose to employ deep reinforcement learning to discard misleading and confounding frames and find the most representative frames from video pairs. With the advantage of deep reinforcement learning, our method can learn strategies to select the optimal frame groups. Experiments show that the proposed framework outperforms the existing methods on the iLIDS-VID, PRID-2011 and MARS datasets.