Federated learning (FL) provides a privacy-preserving solution for distributed machine tasks. One challenging problem that severely damages the performance of FL models is co-occurrence data heterogeneity and long-tail distribution, which frequently appears in real applications. In this paper, we reveal an intriguing fact biased classifier primary factor leading to poor global model. Motivated by above finding, propose novel method heterogeneous long-tailed via Classifier Re-training with...

Data heterogeneity is an inherent challenge that hinders the performance of federated learning (FL). Recent studies have identified biased classifiers local models as key bottleneck. Previous attempts used classifier calibration after FL training, but this approach falls short in improving poor feature representations caused by training-time biases. Resolving bias dilemma requires a full understanding mechanisms behind classifier. advances neural collapse shown and prototypes under perfect...

Federated learning provides a privacy guarantee for generating good deep models on distributed clients with different kinds of data. Nevertheless, dealing non-IID data is one the most challenging problems federated learning. Researchers have proposed variety methods to eliminate negative influence non-IIDness. However, they only focus provided that universal class distribution balanced. In many real-world applications, long-tailed, which causes model seriously biased. Therefore, this paper...

In federated learning (FL), weighted aggregation of local models is conducted to generate a global model, and the weights are normalized (the sum 1) proportional data sizes. this paper, we revisit process gain new insights into training dynamics FL. First, find that can be smaller than 1, causing weight shrinking effect (analogous decay) improving generalization. We explore how optimal factor affected by clients' heterogeneity epochs. Second, dive relative among clients depict importance....

Abstract Greenland Ice Sheet (GrIS) surface melt has contributed to the global sea‐level rise and ongoing warming is expected promote this process. This study provides a new strategy for quantitative estimate of GrIS daily at enhanced resolution (3.125 km) from remote sensing perspective beyond traditional regional climate models (RCMs). Daily flux estimated spaceborne radiometer observations with back‐propagation neural network model. The trained fluxes that are calculated using detailed...

Federated learning (FL) provides a privacy-preserving solution for distributed machine tasks. One challenging problem that severely damages the performance of FL models is co-occurrence data heterogeneity and long-tail distribution, which frequently appears in real applications. In this paper, we reveal an intriguing fact biased classifier primary factor leading to poor global model. Motivated by above finding, propose novel method heterogeneous long-tailed via Classifier Re-training with...

Surface meltwater runoff is believed to be the main cause of alarming mass loss in Greenland Ice Sheet (GrIS); however, recent research has shown that a large amount not directly drained or refrozen but stored form firn aquifers (FAs) interior GrIS. Monitoring changes FAs over GrIS great importance evaluate stability and balance ice sheet. This challenging because are visible on surface direct measurements lacking. A new method proposed map during 2010–2020 period by using C-band Advanced...

Data heterogeneity is an inherent challenge that hinders the performance of federated learning (FL). Recent studies have identified biased classifiers local models as key bottleneck. Previous attempts used classifier calibration after FL training, but this approach falls short in improving poor feature representations caused by training-time biases. Resolving bias dilemma requires a full understanding mechanisms behind classifier. advances neural collapse shown and prototypes under perfect...

The potential impact of increased snowmelt and related hydrological processes on ice sheet stability has become a focus academic attention. Hydro-fracture caused by liquid water is one the main triggers shelf disintegration. Recent discoveries firn aquifer (FA) in Wilkins Ice Shelf (WIS) have updated our understanding surface processes, mass energy balance. However, limited field airborne radar observations FA cannot provide complete picture their distribution characteristics. Microwave...

Personalized Federated Learning (PFL) aims to acquire customized models for each client without disclosing raw data by leveraging the collective knowledge of distributed clients. However, collected in real-world scenarios is likely follow a long-tailed distribution. For example, medical domain, it more common number general health notes be much larger than those specifically relatedto certain diseases. The presence can significantly degrade performance PFL models. Additionally, due diverse...

Federated Semi-Supervised Learning (FSSL) aims to learn a global model from different clients in an environment with both labeled and unlabeled data. Most of the existing FSSL work generally assumes that types data are available on each client. In this paper, we study more general problem setup annotation heterogeneity, where client can hold arbitrary percentage (0%-100%) To end, propose novel framework called Heterogeneously Annotated LEarning (HASSLE). Specifically, it is dual-model two...

The Ice Pathfinder satellite (code: BNU-1) is the first Chinese microsatellite, designed for monitoring polar climate and environmental changes. major payload of BNU-1 wide-field camera which provides multispectral images with a 73.69 m spatial resolution 739 km swath width. However, color misrepresentation issue can be observed as BUN-1 image appears yellowish it gets farther towards center field view (FOV). blue band to higher near FOV declines generously both edge areas image, may cause...

Chinese Herbal Medicine (CHM) is a medicinal treasure of the nation, research on identification and classification CHM can help promote heritage development CHM. In this paper, we propose fast method for common radix rhizome based improved VGG16. Firstly, add Convolutional Block Attention Module (CBAM) to VGG16 improve representation power model. Secondly, reduce number parameters by global average pooling generalization ability The accuracy network model reaches 93.21%, which higher than...

Federated learning provides a privacy guarantee for generating good deep models on distributed clients with different kinds of data. Nevertheless, dealing non-IID data is one the most challenging problems federated learning. Researchers have proposed variety methods to eliminate negative influence non-IIDness. However, they only focus provided that universal class distribution balanced. In many real-world applications, long-tailed, which causes model seriously biased. Therefore, this paper...