Explore the vast range of titles available.

Machine learning methods strive to acquire a robust model during the training process that can effectively generalize to test samples, even in the presence of distribution shifts. However, these methods often suffer from performance degradation due to unknown test distributions. Test-time adaptation (TTA), an emerging paradigm, has the potential to adapt a pre-trained model to unlabeled data during testing, before making predictions. Recent progress in this paradigm has highlighted the significant benefits of using unlabeled data to train self-adapted models prior to inference. In this survey, we categorize TTA into several distinct groups based on the form of test data, namely, test-time domain adaptation, test-time batch adaptation, and online test-time adaptation. For each category, we provide a comprehensive taxonomy of advanced algorithms and discuss various learning scenarios. Furthermore, we analyze relevant applications of TTA and discuss open challenges and promising areas for future research. For a comprehensive list of TTA methods, kindly refer to https://github.com/tim-learn/awesome-test-time-adaptation .

Background Among the most widely predicted climate change-related impacts to biodiversity are geographic range shifts, whereby species shift their spatial distribution to track their climate niches. A series of commonly articulated hypotheses have emerged in the scientific literature suggesting species are expected to shift their distributions to higher latitudes, greater elevations, and deeper depths in response to rising temperatures associated with climate change. Yet, many species are not demonstrating range shifts consistent with these expectations. Here, we evaluate the impact of anthropogenic climate change (specifically, changes in temperature and precipitation) on species’ ranges, and assess whether expected range shifts are supported by the body of empirical evidence. Methods We conducted a Systematic Review, searching online databases and search engines in English. Studies were screened in a two-stage process (title/abstract review, followed by full-text review) to evaluate whether they met a list of eligibility criteria. Data coding, extraction, and study validity assessment was completed by a team of trained reviewers and each entry was validated by at least one secondary reviewer. We used logistic regression models to assess whether the direction of shift supported common range-shift expectations (i.e., shifts to higher latitudes and elevations, and deeper depths). We also estimated the magnitude of shifts for the subset of available range-shift data expressed in distance per time (i.e., km/decade). We accounted for methodological attributes at the study level as potential sources of variation. This allowed us to answer two questions: (1) are most species shifting in the direction we expect (i.e., each observation is assessed as support/fail to support our expectation); and (2) what is the average speed of range shifts? Review findings We found that less than half of all range-shift observations (46.60%) documented shifts towards higher latitudes, higher elevations, and greater marine depths, demonstrating significant variation in the empirical evidence for general range shift expectations. For the subset of studies looking at range shift rates, we found that species demonstrated significant average shifts towards higher latitudes (average = 11.8 km/dec) and higher elevations (average = 9 m/dec), although we failed to find significant evidence for shifts to greater marine depths. We found that methodological factors in individual range-shift studies had a significant impact on the reported direction and magnitude of shifts. Finally, we identified important variation across dimensions of range shifts (e.g., greater support for latitude and elevation shifts than depth), parameters (e.g., leading edge shifts faster than trailing edge for latitude), and taxonomic groups (e.g., faster latitudinal shifts for insects than plants). Conclusions Despite growing evidence that species are shifting their ranges in response to climate change, substantial variation exists in the extent to which definitively empirical observations confirm these expectations. Even though on average, rates of shift show significant movement to higher elevations and latitudes for many taxa, most species are not shifting in expected directions. Variation across dimensions and parameters of range shifts, as well as differences across taxonomic groups and variation driven by methodological factors, should be considered when assessing overall confidence in range-shift hypotheses. In order for managers to effectively plan for species redistribution, we need to better account for and predict which species will shift and by how much. The dataset produced for this analysis can be used for future research to explore additional hypotheses to better understand species range shifts.

Advancement of deep learning algorithms and their application in medical imaging has demonstrated expert-level performance in recent times in connection with diagnosis of different illness of patients. However, their accuracy may drop in the real world as training dataset cannot cover all aspects of distribution. This problem is often encountered when deep neural networks are exposed to out-of-distribution (OOD) instances. This demands the need for employing appropriate technique to handle OOD instances as it is not practically feasible to retrain the models over and over again every time changes occur in distribution of data. In this study, a stack generalization deep ensemble learning approach is proposed for the classification of lung diseases in chest X-ray images. Specifically, this approach includes training a series of M n models on independent D in internal datasets, and then training an advanced meta-model M mlp with the benefits of combined generalization with maximum confidence. Finally, the method has been evaluated with subsets of in-distribution D i n , e x t datasets as well as entirely new OOD dataset D out . The proposed method yields AUC of 0.86, 1.00, and 0.99 for D i n , t e s t test sets. For three entirely new target domain datasets, AUC for D i n , e x t is 0.89, while AUC for D out are found to be 0.86, 0.81, 0.80, and 0.55. This paper highlights the effectiveness of the proposed method in generalization to data distribution shifts in 2D chest X-ray medical imaging data.

When using point-by-point data input with former series models for wind power prediction, the prediction accuracy decreases due to data distribution shifts and the inability to extract local information. To address these issues, this paper proposes an ultra-short-term wind power prediction model based on the Z-score (ZS), Dish-TS (DT), and Patch time series Transformer (PatchTST). Firstly, to reduce the impact of data distribution shift on prediction accuracy, ZS standardization is applied to both training and testing datasets. Additionally, the DT algorithm, which can self-learn the mean and variance, is introduced for window data standardization. Secondly, the PatchTST model is employed to convert point input data into local-level input data. Feature extraction is then performed using the multi-head attention mechanism in the Encoder layer and a feed-forward network composed of one-dimensional convolution to obtain the prediction results. These results are subsequently de-standardized using DT and ZS to restore the original data amplitude. Finally, experimental analysis is conducted, comparing the proposed ZS-DT-PatchTST model with various prediction models. The proposed model achieves the highest prediction accuracy, with a mean absolute error of 5.95 MW, a mean squared error of 10.89 MW, and a coefficient of determination of 97.38%.

Time series forecasting is a critical task that provides key information for decision-making across various fields, such as economic planning, supply chain management, and medical diagnosis. After the use of traditional statistical methodologies and machine learning in the past, various fundamental deep learning architectures such as MLPs, CNNs, RNNs, and GNNs have been developed and applied to solve time series forecasting problems. However, the structural limitations caused by the inductive biases of each deep learning architecture constrained their performance. Transformer models, which excel at handling long-term dependencies, have become significant architectural components for time series forecasting. However, recent research has shown that alternatives such as simple linear layers can outperform Transformers. These findings have opened up new possibilities for using diverse architectures, ranging from fundamental deep learning models to emerging architectures and hybrid approaches. In this context of exploration into various models, the architectural modeling of time series forecasting has now entered a renaissance. This survey not only provides a historical context for time series forecasting but also offers comprehensive and timely analysis of the movement toward architectural diversification. By comparing and re-examining various deep learning models, we uncover new perspectives and present the latest trends in time series forecasting, including the emergence of hybrid models, diffusion models, Mamba models, and foundation models. By focusing on the inherent characteristics of time series data, we also address open challenges that have gained attention in time series forecasting, such as channel dependency, distribution shift, causality, and feature extraction. This survey explores vital elements that can enhance forecasting performance through diverse approaches. These contributions help lower entry barriers for newcomers by providing a systematic understanding of the diverse research areas in time series forecasting (TSF), while offering seasoned researchers broader perspectives and new opportunities through in-depth exploration of TSF challenges.

The use of species distribution models (SDMs) has rapidly increased over the last decade, driven largely by increasing observational evidence of distributional shifts of terrestrial and aquatic populations. These models permit, for example, the quantification of range shifts, the estimation of species co-occurrence, and the association of habitat to species distribution and abundance. The increasing complexity of contemporary SDMs presents new challenges—as the choices among modeling options increase, it is essential to understand how these choices affect model outcomes. Using a combination of original analysis and literature review, we synthesize the effects of three common model choices in semi-parametric predictive process species distribution modeling: model structure, spatial extent of the data, and spatial scale of predictions. To illustrate the effects of these choices, we develop a case study centered around sablefish ( Anoplopoma fimbria ) distribution on the west coast of the USA. The three modeling choices represent decisions necessary in virtually all ecological applications of these methods, and are important because the consequences of these choices impact derived quantities of interest ( e.g ., estimates of population size and their management implications). Truncating the spatial extent of data near the observed range edge, or using a model that is misspecified in terms of covariates and spatial and spatiotemporal fields, led to bias in population biomass trends and mean distribution compared to estimates from models using the full dataset and appropriate model structure. In some cases, these suboptimal modeling decisions may be unavoidable, but understanding the tradeoffs of these choices and impacts on predictions is critical. We illustrate how seemingly small model choices, often made out of necessity or simplicity, can affect scientific advice informing management decisions—potentially leading to erroneous conclusions about changes in abundance or distribution and the precision of such estimates. For example, we show how incorrect decisions could cause overestimation of abundance, which could result in management advice resulting in overfishing. Based on these findings and literature gaps, we outline important frontiers in SDM development.

Accurate runoff forecasting helps mitigate flooding and drought risks and ensure water security under changing conditions. Compared to deterministic prediction models, interval prediction can more effectively quantify uncertainty, enhancing practical applicability. However, the Mixture Density Network (MDN) model—a state‐of‐the‐art probabilistic modeling approach in hydrology—is susceptible to bias from distributional misspecification, and its prediction intervals are often overly wide, reducing practical utility. We therefore innovatively incorporated the Weighted Conformal Inference (WCI) strategy, which accounts for distributional shifts in runoff sequences, and integrated it with MDN to develop the WCI‐MDN model for runoff interval prediction. To validate the effectiveness of the WCI strategy, we constructed six models in total: MDNs and WCI‐MDNs under three distributions—Gaussian Mixture (GMM), Laplace Mixture (LMM), and Countable Mixtures of Asymmetric Laplacians (CMAL)—and evaluated their accuracy and robustness using data from 222 basins in the CAMELS‐AUS data set. Results indicated that among the three MDN models, the LMM distribution achieved the best interval prediction performance, followed by the CMAL and GMM distributions. After introducing the WCI strategy, the coverage width‐based criterion (CWC) for GMM, LMM, and CMAL distributions decreased by approximately 61.1%, 48.7%, and 54.3%, respectively, across all basins, demonstrating that the WCI‐MDNs achieved higher prediction reliability. Furthermore, compared to the MDNs, the standard deviation of the CWC for the WCI‐MDNs was reduced by 66.7%–81.8%, indicating higher robustness. Thus, the study improved the existing MDNs, providing a promising new approach for runoff interval prediction.

Snow cover plays a critical role for many alpine animals, but our understanding of how snow effects vary with elevation is limited. The white‐winged snowfinch is restricted to high‐elevation habitats, relying on snowfields, edges of melting snow patches, and short‐sward alpine grasslands to collect invertebrates for nestlings. To explore potential elevation‐dependent snow effects on breeding occurrence, we monitored 37 nest‐boxes in central Italian Alps from 2017 to 2025 (210 nest‐box–year combinations in total). Nest‐boxes were deployed from 2300 to 3010 m a.s.l., thus encompassing continuous alpine grassland, mosaics of scattered alpine grassland and rocky habitats, and landscapes dominated by rocks and snowfields. Average percentage snow cover was estimated (using remote sensing data) within a 300 m radius around each nest‐box during June (peak of breeding activity), and was used to model nest‐box use at different elevations. June snow cover varied greatly between years (average 5–64%) and influenced nest‐box use in a complex manner. At low elevation (< 2487 m a.s.l.), usage increased with snow cover, whereas it decreased with snow cover at higher elevation (> 2753 m a.s.l.), and was apparently unrelated to snow cover at intermediate ones. At low elevation, increased snow cover means more suitable foraging habitats, including snow patches and the melting margins, as well as short‐sward grassland (the most exploited foraging habitat by breeding snowfinches). At higher elevations, nest‐boxes tended to be occupied more frequently with relatively limited snow cover. Reduced snow cover here implies more suitable short‐sward grassland available to foraging snowfinches, and therefore more profitable conditions. However, above 2900 m a.s.l., nest‐boxes were used only thrice and with high snow cover. There, snowfinches probably rely on wind‐blown arthropods deposited on snowfields, requiring substantial snow cover. The elevation‐dependent patterns we found provide examples of the possible reasons for complex distribution shifts in response to climate change and deserve more investigation.

Accurate cuffless blood pressure (BP) estimation remains challenging, particularly under dynamic conditions with significant intra-individual BP variations. This study introduces UTransBPNet , a novel, calibration-free model for cuffless BP estimation. It integrates a squeeze-and-excitation-enhanced Unet architecture for short-range feature extraction with a transformer and cross attention module to capture long-range dependencies from high-resolution, multi-channel physiological signals, further refined through an optimized fine-tuning scheme. Comprehensive validations were conducted across multiple dynamic datasets—Dataset_Drink, Dataset_Exercise, and Dataset_MIMIC—in both scenario-specific and cross-scenario settings. Results demonstrate that UTransBPNet outperformed existing models in tracking BP variations under dynamic conditions, achieving individual Pearson’s correlation coefficients of 0.61 ± 0.17 and 0.62 ± 0.13 for systolic BP (SBP) and diastolic BP (DBP) in Dataset_Drink, 0.82 ± 0.11 and 0.72 ± 0.18 in Dataset_Exercise, and low mean absolute differences (MADs) of 4.38 and 2.25 mmHg in Dataset_MIMIC. The analysis also highlights the impact of dataset characteristics on model performance, such as distribution shift, distribution imbalance and individual BP variability, highlighting the need for well-curated data to ensure generalizability. This study advances the development of robust, cuffless BP estimation models for real-world applications.

We present a federated learning framework that is designed to robustly deliver good predictive performance across individual clients with heterogeneous data. The proposed approach hinges upon a superquantile-based learning objective that captures the tail statistics of the error distribution over heterogeneous clients. We present a stochastic training algorithm that interleaves differentially private client filtering with federated averaging steps. We prove finite time convergence guarantees for the algorithm: O ( 1 / T ) in the nonconvex case in T communication rounds and O ( exp ( - T / κ 3 / 2 ) + κ / T ) in the strongly convex case with local condition number κ . Experimental results on benchmark datasets for federated learning demonstrate that our approach is competitive with classical ones in terms of average error and outperforms them in terms of tail statistics of the error.