Dynamic source domain selection: An adaptive EEG transfer learning framework for mitigating negative transfer.

The Cancer dependency maps (DepMap) identify genetic dependencies in cancer cells using large-scale loss-of-function screens, providing a foundation for cancer-specific treatment strategies. However, discrepancies exist between cancer cell line models (CCLs) and patient-derived tumor models, particularly in translating findings to clinical settings. To bridge this gap, computational approaches such as artificial intelligence-based domain adaptation can assist in aligning laboratory and patient-derived molecular data, thereby improving the translation of preclinical findings into personalized treatment strategies. We developed a deep unsupervised domain adaptation (UDA) algorithm to align features between source and target domains. It was trained on labeled CCLs data from the source domain and unseen, unlabeled CCL data from the target domain. The trained model was applied to predict the dependency map of breast cancer (BC) patients in The Cancer Genome Atlas (TCGA). To validate its performance, we used the predicted BC dependency map to classify ER + /HER2 + BC subtype statuses and identify synthetic lethality (SL) gene pairs for drug discovery. Our model demonstrated high accuracy in predicting cancer dependency maps for patient-derived tumors. The generated maps showed excellent performance in predicting ER + /HER2+ subtype statuses, with an area under the curve of the receiver operating characteristic (AUC-ROC) of 0.99. Notably, our analysis also identified two potential synthetic lethality gene pairs: PBRM1-NF2 and PBRM1-CTNND2, which can be potentially used for developing precision therapies for ER + /HER2+ breast cancer. Domain adaptation is a promising approach for transferring biological knowledge between different cancer models and improving patient-specific treatment strategies.

LADA: A label-aware framework for cross-domain sentiment classification.

Maximum mean discrepancy enhanced Informer for accurate cross-domain dual-timescale effluent prediction in wastewater treatment plants.

Unsupervised Domain Adaptation (UDA) focuses on transferring knowledge from a labeled source domain to an unlabeled target domain, addressing the challenge of domain shift. Significant domain shifts hinder effective knowledge transfer, leading to negative transfer and deteriorating model performance. Therefore, mitigating negative transfer is essential. This study revisits negative transfer through the lens of causally disentangled learning, emphasizing cross-domain discriminative disagreement on non-causal environmental features as a critical factor. Our theoretical analysis reveals that overreliance on non-causal environmental features as the environment evolves can cause discriminative disagreements (termed environmental disagreement), thereby resulting in negative transfer. To address this, we propose Reducing Environmental Disagreement (RED), which disentangles each sample into domain-invariant causal features and domain-specific non-causal environmental features via adversarially training domain-specific environmental feature extractors in the opposite domains. Subsequently, RED estimates and reduces environmental disagreement based on domain-specific non-causal environmental features. Experimental results confirm that RED effectively mitigates negative transfer and achieves state-of-the-art performance.

Domain adaptive person re-identification with spatiotemporal fusion towards real-world sparse surveillance

Open-set domain adaptation (OSDA) aims to bridge the gap between labeled source and unlabeled target domains while separating unknown data in the target domain. Recent works have addressed the OSDA setting with notable results, yet they have explored inflexible and limited patterns in latent features. The insufficient generalizability of representation in low-density regions against the large variety of unknown data results in the misclassification of sensitive samples during direct alignment. This study proposes a novel strategy that adversarially leverages residual knowledge generalization (ARKG) at the pixel level to ensure boundary consistency for weighted domain alignment and resilient decision-making. To generate distinctive and generalizable source latent features for alignment with the target domain, a new uncertainty-aware residual space (UARS) is produced using a deep residual network influenced by the target domain. In the framework of integrating the variational autoencoder and generative adversarial network (VAE-GAN), source-like images are generated using diverse sampling from this space. Using the Lipschitz continuous principle, a weighting approach is applied to the target domain to maximize the conditional mutual information (MI) between the shared instances and their latent spaces. Finally, hierarchical decisions at the pixel and feature levels draw a dynamic adversarial boundary between known and unknown data. Extensive results on Office-31, Office-Home, DomainNet, and VisDA datasets show that ARKG achieves superior performance by providing state-of-the-art insights in OSDA. Code is available in https://github.com/ReyhaneGhaffari/ARKG.

Accurate kidney ultrasound segmentation is fundamental for clinical measurement and computer-aided diagnosis. However, domain shifts across devices and centers-manifested as differences in grayscale intensity, contrast, and speckle texture statistics-can substantially degrade model generalization, while acquiring new pixel-level annotations is costly. To address this, we propose a statistical spectral-similarity-guided ultrasound-to-ultrasound translation method to improve kidney segmentation performance without target-domain annotations. Motivated by frequency-domain analysis of renal ultrasound data, we observe that mid-to-low frequency components, which encode global organ structure, exhibit high consistency across domains, whereas mid-to-high frequency components, dominated by device-dependent speckle and texture statistics, vary substantially. Based on dataset-level frequency statistics, our method automatically identifies spectrally similar frequency bands shared by the source and target domains and derives structural guidance from them. This guidance is injected as a soft condition throughout a diffusion-based image generation process, enabling translation to target-device appearance while preserving anatomical structure. The translated images, paired with source-domain labels, are then used to train a segmentation network without requiring any target-domain annotations. Experiments on two public renal ultrasound datasets (OKUS and UNK) and an in-house multi-center dataset demonstrate superior structural preservation in image translation and consistently improved downstream segmentation performance, with particularly large reductions in boundary error. In the challenging OKUS to UNK adaptation scenario, our method boosts the mean Dice score by up to 20.52% (from 56.05% to 76.57%) and drastically reduces the 95% Hausdorff Distance (HD95) boundary error by 71.96mm compared to the direct transfer baseline. Furthermore, consistent performance gains are achieved across the in-house multi-center dataset. These results indicate that the proposed spectral-similarity-based guidance effectively handles ultrasound domain shifts, substantially improving robustness and generalization for kidney segmentation under zero-shot and cross-center settings.

Voxel-level annotation for volumetric medical imaging is expensive and difficult to scale, which makes training highcapacity 3-D segmentation models challenging in practice. Transfer learning (TL) from large public datasets is a common remedy, but it can under-perform when the source domain differs from the target anatomy and acquisition characteristics, as is often the case for pulmonary nodules. In this work, we propose a masked autoencoder (MAE) pretraining-based approach to break the data efficiency wall of domain difference and present a focused empirical study of domain-specific self-supervised learning (SSL) for 3-D lung nodule segmentation. We evaluate two experimental settings: first, Masked Autoencoder (MAE) pretraining versus random initialization across representative baselines; second, MAE versus Decathlon TL for UNETR++ while testing whether MAE-based pretraining also benefits a CNN baseline (V-Net). MAE pretraining on target-domain CT volumes achieves a Dice Similarity Coefficient (DSC) of 0.307, outperforming random initialization (0.136) and Decathlon weights (0.257). In addition, MAE improves the stability of V-Net in a "low-data" regime (i.e., with "insufficiently labeled" data), increasing DSC from 0.010 to0.071. Overall, these results suggest that MAE-based pretraining can provide a practical and robust initialization strategy for volumetric segmentation when labeled data are limited.

Theories of metaphorical embodiment have widely treated perceptual processes as an inseparable component of metaphoric conceptualization. This paper critically examines and synthesizes influential accounts of metaphor in cognitive linguistics and cognitive science, with particular attention to Gentner's structure-mapping theory, Lakoff's invariance principle, and Ruiz de Mendoza Ibáñez's extended invariance principle. We propose a distinction between metaphoric conceptualization and metaphoric perceptualization as analytically separable, though interrelated, processes. Our discussion is limited to metaphors that are interpreted through the projection of image-schematic or propositional structures. We suggest that metaphoric conceptualization primarily involves structural projection, whereby the abstract structure of a source domain-independent of its perceivable features-is mapped onto a target domain. Metaphoric perceptualization, by contrast, consists in the organization of the target domain's perceivable features within this projected structure, as well as, in some cases, the attribution of salient perceivable features from the source domain to the target domain. Through this process, a novel perceptual representation of the target domain may emerge. By integrating and reassessing existing theoretical frameworks, we conclude that metaphorical embodiment is best understood as comprising two distinct but related processes, with metaphoric perceptualization being guided and constrained by metaphoric conceptualization.

Open set domain adaptation (OSDA) faces two critical challenges: the emergence of unknown classes in the target domain and changes in observed distributions across domains. Although numerous studies have proposed advanced algorithms, recent experimental results demonstrate that the classical empirical risk minimization (ERM) approach still delivers state-of-the-art performance. However, few theories can effectively explain this disputed phenomenon. To address the theoretical gap, we focus on constructing a causal theoretical framework for OSDA. We formulate the novel concepts of the fully informative causal invariance model (FICIM) and the partially informative causal invariance model (PICIM). Subsequently, we derive an OSDA theoretical bound to prove that the ERM performs well when the source domain follows FICIM, while it performs poorly when the source domain follows PICIM. The different results may be attributed to the varying amounts of available information when bounding the target domain's stable expected risk. Finally, across different datasets, we conduct extensive experiments on the FICIM and PICIM source domains to validate the effectiveness of our theoretical results. Moreover, our findings can also support the training and fine-tuning of large language models (LLMs).

Domain shift, characterized by degraded model performance during the transfer from labeled source domains to unlabeled target domains, poses a persistent challenge for deploying deep learning systems. Current unsupervised domain adaptation (UDA) methods predominantly rely on fine-tuning feature extractors-an approach limited by high computational cost, reduced interpretability, and poor scalability to modern architectures. Our analysis reveals that models pre-trained on large-scale data exhibit domain-invariant geometric patterns in their feature space, characterized by intra-class clustering and inter-class separation, thereby preserving transferable discriminative structures. These findings suggest that cross-domain performance degradation is often associated with decision-boundary misalignment, and that correcting such misalignment can serve as an effective alternative to feature adaptation, particularly when pretrained representations are sufficiently strong. Unlike fine-tuning entire pre-trained models, which risks introducing unpredictable feature distortions, we propose the Feature-space Planes Searcher (FPS): a novel domain adaptation framework that optimizes decision boundaries by leveraging these geometric patterns while keeping the feature encoder frozen. This streamlined approach enables interpretable analysis of adaptation while substantially reducing memory and computational costs through offline feature extraction, permitting full-dataset optimization in a single training cycle. Moreover, we introduce an Intra-Class Distance Metric (ICDM) that enables fully unsupervised hyperparameter selection without requiring target-domain labels. Evaluations on public benchmarks show that FPS achieves competitive performance across standard benchmarks, with notable gains in several settings and tasks. FPS scales efficiently with large multimodal models and shows versatility across diverse domains including protein structure prediction, remote sensing classification, and earthquake detection. We anticipate FPS will provide a simple, effective, and generalizable framework for domain adaptation tasks.

Application of instance-based transfer learning for quantitative detection of honey adulteration.

Learnable wavelet packet kernel guided deep discriminative dynamic joint domain adaptation network for cross-machine fault diagnosis under strong noise.

Glioma, a highly invasive tumor of the central nervous system with poor prognosis, requires accurate detection for effective clinical management, where rapid and precise tissue discrimination plays a critical role. Raman spectroscopy shows strong potential for real-time detection; however, spectral variations across different systems make it difficult to directly apply models trained on a master instrument (source domain) to a slave instrument (target domain). Therefore, model transfer becomes a key challenge. Existing methods typically rely on transfer set samples, requiring paired measurements of the same samples on both instruments to establish a mapping relationship, which is often impractical in real-world scenarios. Moreover, the limited number of target domain samples makes direct modeling prone to overfitting. To address these challenges, this paper proposes a Subdomain Feature Alignment Network (SFAN). Instead of performing spectral mapping, the proposed method conducts class-conditional alignment in the feature space by minimizing the Local Maximum Mean Discrepancy (LMMD), thereby learning domain-invariant and discriminative feature representations. To improve transfer stability under small sample conditions, a collaborative soft-hard label weighting mechanism is designed. Hard labels are used to guide the alignment direction, while soft labels are introduced to capture the probabilistic structure within classes, reducing the risk of incorrect alignment and alleviating overfitting. In addition, a two-stage network migration strategy is proposed to decouple cross-domain shared feature learning from target domain adaptation. This allows the model to first learn stable and generalizable features, followed by fine-tuning on the target domain, thereby enhancing transfer performance and robustness. Experimental results demonstrate that the proposed method outperforms conventional approaches on a human glioma dataset. By shifting from spectral mapping to feature alignment, the proposed method fundamentally improves model transferability and can be applied across different instruments and under varying measurement conditions. It provides a more general and effective framework for cross-domain modeling of Raman spectroscopy analysis under small sample scenarios.

Context-aware reliability exploration for unsupervised domain adaptive person re-identification.

Leveraging VLMs for MUDA: Category-specific prompt with multi-modal interactive LoRA.

Smart perception systems are essential for detecting complex physical and chemical stimuli in diverse environmental monitoring and clinical diagnostic applications. However, the escalating demands for multi-functional inference, cross-scenario deployment, and long-term stability remain difficult to satisfy simultaneously within existing sensing frameworks. This work proposes a unified and computationally efficient deep-learning framework that integrates multi-task learning, transfer learning, and domain adaptation under a shared backbone to resolve these fragmented reliability bottlenecks. Using gas sensing as a representative modality, a lightweight, task-aligned model is developed to concurrently predict sensor working status, gas identity, and gas concentration from transient responses while maintaining a minimal parameter footprint. To bridge the gap between black-box decision logic and physical sensing mechanisms, SHapley Additive exPlanations (SHAP) analysis is employed to quantify multi-scale attributions, elucidate multi-task synergy, and guide sensor-array lightweighting. For cross-scenario scalability, a few-shot structural transfer strategy utilizing parameter-efficient fine-tuning is introduced to facilitate rapid adaptation to heterogeneous domains. To ensure cross-period robustness under baseline drift, a semi-supervised adversarial domain-adaptation scheme with dual statistical alignment is implemented to mitigate distribution shifts. Across diverse datasets, the framework achieves high accuracy (>0.98 in the source domain and >0.91 in adaptation settings) with minimal fine-tuning overhead (trainable parameters <2%) and significantly enhanced robustness against sensor drift (up to 24.7% gain). This work provides an interpretable and resource-efficient methodological foundation for deployable intelligent sensing systems, enabling cohesive cross-task, cross-scenario, and cross-period reliability.

ABSTRACT Discussions of proposed climate solutions, such as carbon dioxide removal (CDR), are multi-layered and contested. This study examines the role that metaphors play as frame devices in news coverage (2018–2024) about CDR. Using critical metaphor analysis, we examined 257 articles from major UK, US, and Canadian news outlets to identify and interpret contrasting metaphorical expressions from journalists and their sources, including industry, science, and civil society. We find that a wide range of source domains, including references to, e.g. historical events, household objects, crime, religion, and medical analogies, is used to metaphorically frame CDR. These metaphors reflect actors’ competing ideologies and interests, rooted in hopeful rational-optimist and socio-ecological visions. We also discuss how metaphor use could influence public engagement and policy and reflect on how language might oversimplify or obscure critical aspects of the technology.

Integration of Geostatistical and Hot Spot Analysis to Identify Drivers of Radionuclide Spatial Heterogeneity in the High-Mountain Lake Basin Topsoil.