Cross-domain Setting Research Articles

EntroCap: Zero-shot image captioning with entropy-based retrieval

State-of-the-art image captioning often relies on supervised models trained on domain-specific image–text pairs, which can be expensive and time-consuming to annotate. Additionally, these models struggle with performance drops when applied to data from different distributions. To overcome these challenges, zero-shot image captioning has emerged, utilizing models trained on unlabeled image and text data. Pre-trained vision-language and generative language models have shown promise in this area, but existing methods do not fully leverage the cross-modal representation capabilities of large-scale models and tend to overlook smaller objects in images. We propose EntroCap, a novel zero-shot image captioning method that incorporates an Entropy-based Retrieval Strategy. During training, our Hierarchical Projector optimizes CLIP’s joint embeddings to capture comprehensive global signals with multi-level representations. During inference, the Entropy-based Retrieval Strategy adjusts prediction logits to better include smaller, less frequent objects. Additionally, our Balancing Gate mechanism guides the language model to produce more detailed and contextually relevant captions by balancing local and global signals. Extensive experiments on MSCOCO, Flickr30k, and NoCaps demonstrate that EntroCap outperforms previous zero-shot approaches in both in-domain and cross-domain settings. We also conduct experiments to quantitatively validate EntroCap’s effectiveness in describing small objects.

MMDG-DTI: Drug–target interaction prediction via multimodal feature fusion and domain generalization

Recently, deep learning has become the essential methodology for Drug–Target Interaction (DTI) prediction. However, the existing learning-based representation methods embed the prior knowledge encapsulated by the training data and, as such, acquire redundant domain information irrelevant to the DTI prediction, compromising the generalization capability of a deep network to unfamiliar instances. To tackle this limitation, we propose a novel DTI prediction method using Multi-Modal feature fusion and Domain Generalization (MMDG-DTI). Specifically, we first use pre-trained Large Language Models (LLMs) to obtain generalized textual features covering nearly the whole biological text vocabulary, with the capacity to handle unseen samples and extract robust and discriminative features. In parallel, we propose a unified structural feature extractor based on a hybrid Graph Neural Network (GNN). The extractor improves the performance of MMDG-DTI by extracting features from another modality and discarding the representation of domain-specific prior substructures. Then, we integrate the complementary information of LLM and GNN features using an innovative classifier, while enhancing the generalization ability with the help of Domain Adversarial Training (DAT) and contrastive learning. Last, we propose a novel cross-domain evaluation protocol to enhance the predication fidelity in practical applications. The quantitative and visualization results obtained on several benchmarking datasets demonstrate that MMDG-DTI achieves state-of-the-art performance under both single-domain and cross-domain settings, confirming the superiority of the proposed method. The datasets and codes will be available on the project homepage: https://github.com/JU-HuaY/MMDG-DTI.

Domain-consistent syntactic representation for cross-domain aspect sentiment triplet extraction

Aspect Sentiment Triplet Extraction (ASTE) aims to identify triplets consisting of aspect terms, opinion terms, and sentiment polarities in reviews. Although previous methods of training specific models for a single domain have achieved refreshing results, the assumption that there are enough labeled samples in each domain does not always hold, and the performance of most models in domains lacking labeled samples can suffer significant degradation. To this end, we explore the ASTE task under a cross-domain setting to solve the problem of sentiment triplet extraction in unlabeled sample scenarios. In this paper, we propose a domain-consistent syntactic representation method, which aims to transfer the knowledge mined from resource-rich source domains to unlabeled target domains by minimizing the discrepancy of syntactic representations between domains. Specifically, we design a part-of-speech perception network that captures the part-of-speech rules of words playing the role of task-related entities and a dependency perception network that learns potential dependencies between words. Meanwhile, Maximum Mean Discrepancy (MMD) is adopted to minimize the discrepancy of syntactic representations containing rule information between domains to obtain domain-consistent syntactic representation, thereby transferring mined knowledge and providing domain-invariant features for classifier decision-making. Extensive experimental results on eight different domain datasets demonstrate the superiority of our method.

Versatile Framework for Drug-Target Interaction Prediction by Considering Domain-Specific Features.

Predicting drug-target interactions (DTIs) is one of the crucial tasks in drug discovery, but traditional wet-lab experiments are costly and time-consuming. Recently, deep learning has emerged as a promising tool for accelerating DTI prediction due to its powerful performance. However, the models trained on limited known DTI data struggle to generalize effectively to novel drug-target pairs. In this work, we propose a strategy to train an ensemble of models by capturing both domain-generic and domain-specific features (E-DIS) to learn diverse domain features and adapt them to out-of-distribution data. Multiple experts were trained on different domains to capture and align domain-specific information from various distributions without accessing any data from unseen domains. E-DIS provides a comprehensive representation of proteins and ligands by capturing diverse features. Experimental results on four benchmark data sets in both in-domain and cross-domain settings demonstrated that E-DIS significantly improved model performance and domain generalization compared to existing methods. Our approach presents a significant advancement in DTI prediction by combining domain-generic and domain-specific features, enhancing the generalization ability of the DTI prediction model.

Aligning the domains in cross domain model inversion attack

Model Inversion Attack reconstructs confidential training dataset from a target deep learning model. Most of the existing methods assume the adversary has an auxiliary dataset that has similar distribution with the private dataset. However, this assumption does not always hold in real-world scenarios. Since the private dataset is unknown, the domain divergence between the auxiliary dataset and the private dataset is inevitable. In this paper, we use Cross Domain Model Inversion Attack to represent the distribution divergence scenario in MIA. With the distribution divergence between the private images and auxiliary images, the distribution between the feature vectors of the private images and those of the auxiliary images is also different. Moreover, the outputted prediction vectors of the auxiliary images are also misclassified. The inversion attack is thus hard to be performed. We perform both the feature vector inversion task and prediction vector inversion task in this cross domain setting. For feature vector inversion, Domain Alignment MIA (DA-MIA) is proposed. While performing the reconstruction task, DA-MIA aligns the feature vectors of auxiliary images with the feature vectors of private images in an adversarial manner to mitigate the domain divergence between them. Thus, semantically meaningful images can be reconstructed. For prediction vector inversion, we further introduce an auxiliary classifier and propose Domain Alignment MIA with Auxiliary Classifier (DA-MIA-AC). The auxiliary classifier is pretrained by the auxiliary dataset and fine-tuned during the adversarial training stage. Thus, the misclassification problem caused by domain divergence can be solved, and the images can be reconstructed correctly. Various experiments are performed to show the advancement of our methods, the results show that DA-MIA can improve the SSIM score of the reconstructed images for up to 191%, DA-MIA-AC can increase the classification accuracy score of the reconstructed images from 9.18% to 81.32% in Cross Domain Model Inversion Attack.

SemTra: A Semantic Skill Translator for Cross-Domain Zero-Shot Policy Adaptation

This work explores the zero-shot adaptation capability of semantic skills, semantically interpretable experts' behavior patterns, in cross-domain settings, where a user input in interleaved multi-modal snippets can prompt a new long-horizon task for different domains. In these cross-domain settings, we present a semantic skill translator framework SemTra which utilizes a set of multi-modal models to extract skills from the snippets, and leverages the reasoning capabilities of a pretrained language model to adapt these extracted skills to the target domain. The framework employs a two-level hierarchy for adaptation: task adaptation and skill adaptation. During task adaptation, seq-to-seq translation by the language model transforms the extracted skills into a semantic skill sequence, which is tailored to fit the cross-domain contexts. Skill adaptation focuses on optimizing each semantic skill for the target domain context, through parametric instantiations that are facilitated by language prompting and contrastive learning-based context inferences. This hierarchical adaptation empowers the framework to not only infer a complex task specification in one-shot from the interleaved multi-modal snippets, but also adapt it to new domains with zero-shot learning abilities. We evaluate our framework with Meta-World, Franka Kitchen, RLBench, and CARLA environments. The results clarify the framework's superiority in performing long-horizon tasks and adapting to different domains, showing its broad applicability in practical use cases, such as cognitive robots interpreting abstract instructions and autonomous vehicles operating under varied configurations.

Image Captioning with Multi-Context Synthetic Data

Image captioning requires numerous annotated image-text pairs, resulting in substantial annotation costs. Recently, large models (e.g. diffusion models and large language models) have excelled in producing high-quality images and text. This potential can be harnessed to create synthetic image-text pairs for training captioning models. Synthetic data can improve cost and time efficiency in data collection, allow for customization to specific domains, bootstrap generalization capability for zero-shot performance, and circumvent privacy concerns associated with real-world data. However, existing methods struggle to attain satisfactory performance solely through synthetic data. We identify the issue as generated images from simple descriptions mostly capture a solitary perspective with limited context, failing to align with the intricate scenes prevalent in real-world imagery. To tackle this, we present an innovative pipeline that introduces multi-context data generation. Beginning with an initial text corpus, our approach employs a large language model to extract multiple sentences portraying the same scene from diverse viewpoints. These sentences are then condensed into a single sentence with multiple contexts. Subsequently, we generate intricate images using the condensed captions through diffusion models. Our model is exclusively trained on synthetic image-text pairs crafted through this process. The effectiveness of our pipeline is validated through experimental results in both the in-domain and cross-domain settings, where it achieves state-of-the-art performance on well-known datasets such as MSCOCO, Flickr30k, and NoCaps.

Multilevel Attention Network with Semi-supervised Domain Adaptation for Drug-Target Prediction

Prediction of drug-target interactions (DTIs) is a crucial step in drug discovery, and deep learning methods have shown great promise on various DTI datasets. However, existing approaches still face several challenges, including limited labeled data, hidden bias issue, and a lack of generalization ability to out-of-domain data. These challenges hinder the model's capacity to learn truly informative interaction features, leading to shortcut learning and inferior predictive performance on novel drug-target pairs. To address these issues, we propose MlanDTI, a semi-supervised domain adaptive multilevel attention network (Mlan) for DTI prediction. We utilize two pre-trained BERT models to acquire bidirectional representations enriched with information from unlabeled data. Then, we introduce a multilevel attention mechanism, enabling the model to learn domain-invariant DTIs at different hierarchical levels. Moreover, we present a simple yet effective semi-supervised pseudo-labeling method to further enhance our model's predictive ability in cross-domain scenarios. Experiments on four datasets show that MlanDTI achieves state-of-the-art performances over other methods under intra-domain settings and outperforms all other approaches under cross-domain settings. The source code is available at https://github.com/CMACH508/MlanDTI.

Transferable Video Moment Localization by Moment-Guided Query Prompting

Video moment localization stands as a crucial task within the realm of computer vision, entailing the identification of temporal moments in untrimmed videos that bear semantic relevance to the supplied natural language queries. This work delves into a relatively unexplored facet of the task: the transferability of video moment localization models. This concern is addressed by evaluating moment localization models within a cross-domain transfer setting. In this setup, we curate multiple datasets distinguished by substantial domain gaps. The model undergoes training on one of these datasets, while validation and testing are executed using the remaining datasets. To confront the challenges inherent in this scenario, we draw inspiration from the recently introduced large-scale pre-trained vision-language models. Our focus is on exploring how the strategic utilization of these resources can bolster the capabilities of a model designed for video moment localization. Nevertheless, the distribution of language queries in video moment localization usually diverges from the text used by pre-trained models, exhibiting distinctions in aspects such as length, content, expression, and more. To mitigate the gap, this work proposes a Moment-Guided Query Prompting (MGQP) method for video moment localization. Our key idea is to generate multiple distinct and complementary prompt primitives through stratification of the original queries. Our approach is comprised of a prompt primitive constructor, a multimodal prompt refiner, and a holistic prompt incorporator. We carry out extensive experiments on Charades-STA, TACoS, DiDeMo, and YouCookII datasets, and investigate the efficacy of the proposed method using various pre-trained models, such as CLIP, ActionCLIP, CLIP4Clip, and VideoCLIP. The experimental results demonstrate the effectiveness of our proposed method.

FACL-Attack: Frequency-Aware Contrastive Learning for Transferable Adversarial Attacks

Deep neural networks are known to be vulnerable to security risks due to the inherent transferable nature of adversarial examples. Despite the success of recent generative model-based attacks demonstrating strong transferability, it still remains a challenge to design an efficient attack strategy in a real-world strict black-box setting, where both the target domain and model architectures are unknown. In this paper, we seek to explore a feature contrastive approach in the frequency domain to generate adversarial examples that are robust in both cross-domain and cross-model settings. With that goal in mind, we propose two modules that are only employed during the training phase: a Frequency-Aware Domain Randomization (FADR) module to randomize domain-variant low- and high-range frequency components and a Frequency-Augmented Contrastive Learning (FACL) module to effectively separate domain-invariant mid-frequency features of clean and perturbed image. We demonstrate strong transferability of our generated adversarial perturbations through extensive cross-domain and cross-model experiments, while keeping the inference time complexity.

Segmentation-based cardiomegaly detection based on semi-supervised estimation of cardiothoracic ratio

The successful integration of neural networks in a clinical setting is still uncommon despite major successes achieved by artificial intelligence in other domains. This is mainly due to the black box characteristic of most optimized models and the undetermined generalization ability of the trained architectures. The current work tackles both issues in the radiology domain by focusing on developing an effective and interpretable cardiomegaly detection architecture based on segmentation models. The architecture consists of two distinct neural networks performing the segmentation of both cardiac and thoracic areas of a radiograph. The respective segmentation outputs are subsequently used to estimate the cardiothoracic ratio, and the corresponding radiograph is classified as a case of cardiomegaly based on a given threshold. Due to the scarcity of pixel-level labeled chest radiographs, both segmentation models are optimized in a semi-supervised manner. This results in a significant reduction in the costs of manual annotation. The resulting segmentation outputs significantly improve the interpretability of the architecture’s final classification results. The generalization ability of the architecture is assessed in a cross-domain setting. The assessment shows the effectiveness of the semi-supervised optimization of the segmentation models and the robustness of the ensuing classification architecture.

Triple Sequence Learning for Cross-domain Recommendation

Cross-domain recommendation (CDR) aims at leveraging the correlation of users’ behaviors in both the source and target domains to improve the user preference modeling in the target domain. Conventional CDR methods typically explore the dual-relations between the source and target domains’ behaviors. However, this may ignore the informative mixed behaviors that naturally reflect the user’s global preference. To address this issue, we present a novel framework, termed triple sequence learning for cross-domain recommendation (Tri-CDR), which jointly models the source, target, and mixed behavior sequences to highlight the global and target preference and precisely model the triple correlation in CDR. Specifically, Tri-CDR independently models the hidden representations for the triple behavior sequences and proposes a triple cross-domain attention (TCA) method to emphasize the informative knowledge related to both user’s global and target-domain preference. To comprehensively explore the cross-domain correlations, we design a triple contrastive learning (TCL) strategy that simultaneously considers the coarse-grained similarities and fine-grained distinctions among the triple sequences, ensuring the alignment while preserving information diversity in multi-domain. We conduct extensive experiments and analyses on six cross-domain settings. The significant improvements of Tri-CDR with different sequential encoders verify its effectiveness and universality. The source code is available at https://github.com/hulkima/Tri-CDR .

An Emotion-Aware Multitask Approach to Fake News and Rumor Detection Using Transfer Learning

Social networking sites, blogs, and online articles are instant sources of news for internet users globally. However, in the absence of strict regulations mandating the genuineness of every text on social media, it is probable that some of these texts are fake news or rumors. Their deceptive nature and ability to propagate instantly can have an adverse effect on society. This necessitates the need for more effective detection of fake news and rumors on the web. In this work, we annotate four fake news detection (EFN) and rumor detection datasets with their emotion class labels using transfer learning. We show the correlation between the legitimacy of a text with its intrinsic emotion for fake news and rumor detection, and prove that even within the same emotion class, fake and real news are often represented differently, which can be used for improved feature extraction. Based on this, we propose a multitask framework for fake news and rumor detection, predicting both the emotion and legitimacy of the text. We train a variety of deep learning models in single-task (STL) and multitask settings for a more comprehensive comparison. We further analyze the performance of our multitask approach for EFN in cross-domain settings to verify its efficacy for better generalization across datasets, and to verify that emotions act as a domain-independent feature. Experimental results verify that our multitask models consistently outperform their STL counterparts in terms of accuracy, precision, recall, and F1 score, both for in-domain and cross-domain settings. We also qualitatively analyze the difference in performance in STL and multitask learning (MTL) models.

Improving Text-Based Person Retrieval by Excavating All-Round Information Beyond Color.

Text-based person retrieval is the process of searching a massive visual resource library for images of a particular pedestrian, based on a textual query. Existing approaches often suffer from a problem of color (CLR) over-reliance, which can result in a suboptimal person retrieval performance by distracting the model from other important visual cues such as texture and structure information. To handle this problem, we propose a novel framework to Excavate All-round Information Beyond Color for the task of text-based person retrieval, which is therefore termed EAIBC. The EAIBC architecture includes four branches, namely an RGB branch, a grayscale (GRS) branch, a high-frequency (HFQ) branch, and a CLR branch. Furthermore, we introduce a mutual learning (ML) mechanism to facilitate communication and learning among the branches, enabling them to take full advantage of all-round information in an effective and balanced manner. We evaluate the proposed method on three benchmark datasets, including CUHK-PEDES, ICFG-PEDES, and RSTPReid. The experimental results demonstrate that EAIBC significantly outperforms existing methods and achieves state-of-the-art (SOTA) performance in supervised, weakly supervised, and cross-domain settings.

Contrastive Graph Prompt-tuning for Cross-domain Recommendation

Recommender systems commonly suffer from the long-standing data sparsity problem where insufficient user-item interaction data limits the systems’ ability to make accurate recommendations. This problem can be alleviated using cross-domain recommendation techniques. In particular, in a cross-domain setting, knowledge sharing between domains permits improved effectiveness on the target domain. While recent cross-domain recommendation techniques used a pre-training configuration, we argue that such techniques lead to a low fine-tuning efficiency, especially when using large neural models. In recent language models, prompts have been used for parameter-efficient and time-efficient tuning of the models on the downstream tasks—these prompts represent a tunable latent vector that permits to freeze the rest of the language model’s parameters. To address the cross-domain recommendation task in an efficient manner, we propose a novel Personalised Graph Prompt-based Recommendation (PGPRec) framework, which leverages the efficiency benefits from prompt-tuning. In such a framework, we develop personalised and item-wise graph prompts based on relevant items to those items the user has interacted with. In particular, we apply Contrastive Learning to generate the pre-trained embeddings, to allow an increased generalisability in the pre-training stage, and to ensure an effective prompt-tuning stage. To evaluate the effectiveness of our PGPRec framework in a cross-domain setting, we conduct an extensive evaluation with the top- k recommendation task and perform a cold-start analysis. The obtained empirical results on four Amazon Review datasets show that our proposed PGPRec framework can reduce up to 74% of the tuned parameters with a competitive performance and achieves an 11.41% improved performance compared to the strongest baseline in a cold-start scenario.

Keyphrase generation for the Russian-language scientific texts using mT5

In this work, we applied the multilingual text-to-text transformer (mT5) to the task of keyphrase generation for Russian scientific texts using the Keyphrases CS&Math Russian corpus. The automatic selection of keyphrases is a relevant task of natural language processing since keyphrases help readers find the article easily and facilitate the systematization of scientific texts. In this paper, the task of keyphrase selection is considered as a text summarization task. The mT5 model was fine-tuned on the texts of abstracts of Russian research papers. We used abstracts as an input of the model and lists of keyphrases separated with commas as an output. The results of mT5 were compared with several baselines, including TopicRank, YAKE!, RuTermExtract, and KeyBERT. The results are reported in terms of the full-match F1-score, ROUGE-1, and BERTScore. The best results on the test set were obtained by mT5 and RuTermExtract. The highest F1-score is demonstrated by mT5 (11,24 %), exceeding RuTermExtract by 0,22 %. RuTermextract shows the highest score for ROUGE-1 (15,12 %). According to BERTScore, the best results were also obtained using these methods: mT5 — 76,89 % (BERTScore using mBERT), RuTermExtract — 75,8 % (BERTScore using ruSciBERT). Moreover, we evaluated the capability of mT5 for predicting the keyphrases that are absent in the source text. The important limitations of the proposed approach are the necessity of having a training sample for fine-tuning and probably limited suitability of the fine-tuned model in cross-domain settings. The advantages of keyphrase generation using pre-trained mT5 are the absence of the need for defining the number and length of keyphrases and normalizing produced keyphrases, which is important for flective languages, and the ability to generate keyphrases that are not presented in the text explicitly.

Omni-Training: Bridging Pre-Training and Meta-Training for Few-Shot Learning.

Few-shot learning aims to fast adapt a deep model from a few examples. While pre-training and meta-training can create deep models powerful for few-shot generalization, we find that pre-training and meta-training focus respectively on cross-domain transferability and cross-task transferability, which restricts their data efficiency in the entangled settings of domain shift and task shift. We thus propose the Omni-Training framework to seamlessly bridge pre-training and meta-training for data-efficient few-shot learning. Our first contribution is a tri-flow Omni-Net architecture. Besides the joint representation flow, Omni-Net introduces two parallel flows for pre-training and meta-training, responsible for improving domain transferability and task transferability respectively. Omni-Net further coordinates the parallel flows by routing their representations via the joint-flow, enabling knowledge transfer across flows. Our second contribution is the Omni-Loss, which introduces a self-distillation strategy separately on the pre-training and meta-training objectives for boosting knowledge transfer throughout different training stages. Omni-Training is a general framework to accommodate many existing algorithms. Evaluations justify that our single framework consistently and clearly outperforms the individual state-of-the-art methods on both cross-task and cross-domain settings in a variety of classification, regression and reinforcement learning problems.

Multi-class point cloud completion networks for 3D cardiac anatomy reconstruction from cine magnetic resonance images.

Cine magnetic resonance imaging (MRI) is the current gold standard for the assessment of cardiac anatomy and function. However, it typically only acquires a set of two-dimensional (2D) slices of the underlying three-dimensional (3D) anatomy of the heart, thus limiting the understanding and analysis of both healthy and pathological cardiac morphology and physiology. In this paper, we propose a novel fully automatic surface reconstruction pipeline capable of reconstructing multi-class 3D cardiac anatomy meshes from raw cine MRI acquisitions. Its key component is a multi-class point cloud completion network (PCCN) capable of correcting both the sparsity and misalignment issues of the 3D reconstruction task in a unified model. We first evaluate the PCCN on a large synthetic dataset of biventricular anatomies and observe Chamfer distances between reconstructed and gold standard anatomies below or similar to the underlying image resolution for multiple levels of slice misalignment. Furthermore, we find a reduction in reconstruction error compared to a benchmark 3D U-Net by 32% and 24% in terms of Hausdorff distance and mean surface distance, respectively. We then apply the PCCN as part of our automated reconstruction pipeline to 1000 subjects from the UK Biobank study in a cross-domain transfer setting and demonstrate its ability to reconstruct accurate and topologically plausible biventricular heart meshes with clinical metrics comparable to the previous literature. Finally, we investigate the robustness of our proposed approach and observe its capacity to successfully handle multiple common outlier conditions.

Dual-model self-regularization and fusion for domain adaptation of robust speaker verification

Learning robust representations of speaker identity is a key challenge in speaker verification, as it results in good generalization for many real-world speaker verification scenarios with domain or intra-speaker variations. In this study, we aim to improve the well-established ECAPA-TDNN framework to enhance its domain robustness for low-resource cross-domain speaker verification tasks. Specifically, a novel dual-model self-learning approach is first proposed to produce robust speaker identity embeddings, where the ECAPA-TDNN is extended into a dual-model structure and then trained and regularized using self-supervised learning between different intermediate acoustic representations; Then, we enhance the dual-model by combining self-supervised loss and supervised loss in a time-dependent manner, thus enhancing the model’s overall generalization capabilities. Furthermore, to better utilize the complementary information in the dual-model’s outputs, we explore various methods for similarity computation and score fusion. Our experiments, conducted on the publicly available VoxCeleb2 and VoxMovies datasets, have demonstrated that our proposed dual-model regularization and fusion methods outperformed the strong baseline by a relative 9.07%–11.6% EER reduction across various in-domain and cross-domain evaluation sets. Importantly, our approach exhibits effectiveness in both supervised and unsupervised scenarios for low-resource cross-domain speaker verification tasks.

Dynamic gradient reactivation for backward compatible person re-identification

We study the backward compatible problem for person re-identification (Re-ID), which aims to constrain the features of an updated new model to be comparable with the existing features from the old model in galleries. Most of the existing works adopt distillation-based methods, which focus on pushing new features to imitate the distribution of the old ones. However, the distillation-based methods are intrinsically sub-optimal since it forces the new feature space to imitate the inferior old feature space. To address this issue, we propose the Ranking-based Backward Compatible Learning (RBCL), which directly optimizes the ranking metric between new features and old features. Different from previous methods, RBCL only pushes the new features to find best-ranking positions in the old feature space instead of strictly alignment, and is in line with the ultimate goal of backward retrieval. However, the sharp sigmoid function used to make the ranking metric differentiable also incurs the gradient vanish issue, therefore stems the ranking refinement during the later period of training. To address this issue, we propose the Dynamic Gradient Reactivation (DGR), which can reactivate the suppressed gradients by adding dynamically computed constants during the forward step. To further help target the best-ranking positions, we include the Neighbor Context Agents (NCAs) to approximate the entire old feature space during training. Unlike previous works that mainly test on the in-domain settings, we make the early attempt to introduce the cross-domain settings (including both supervised and unsupervised) for the backward compatible person Re-ID task, which are more challenging yet meaningful. The experimental results on all five settings show that the proposed RBCL outperforms previous state-of-the-art methods by large margins.