YZ Sözlüğü
Yapay Zekanın tam sözlüğü
Non-IID Data Distribution
Data distribution where samples do not follow an identical and independent distribution across clients, creating statistical heterogeneity that complicates federated learning.
Client Drift
Phenomenon where local client models gradually diverge from the global model due to their non-IID local data distributions, compromising federated learning convergence.
Local Statistical Heterogeneity
Statistical variation of data between different clients, measured by differences in feature and label distributions, constituting the main challenge of non-IID federated learning.
Federated Averaging (FedAvg)
Aggregation algorithm where local model weights are averaged proportionally to each client's data size, but suffers from degraded performance in non-IID contexts.
Personalized Federated Learning
Approach aimed at generating personalized models for each client while benefiting from federated collaboration, essential for effectively managing non-IID data.
Client Selection
Optimal participant selection strategy for each training round to mitigate the negative effects of data heterogeneity and improve overall convergence.
Communication Efficiency
Optimization of exchanges between server and clients in a non-IID context, where more iterations are often needed to achieve convergence compared to the IID case.
Aggregation Strategy
Method of combining local model updates, specifically adapted to handle non-IID data heterogeneity and avoid performance degradation.
Weight Divergence
Measure quantifying the gap between local and global model weights, which increases significantly in the presence of non-IID data distributions.
Feature Heterogeneity
Type of heterogeneity where feature distributions differ between clients, creating distinct local representation spaces that complicate federated aggregation.
Label Distribution Skew
Bias in label distribution where certain clients have over-representation of specific classes, dramatically affecting the global model's performance.
Quantity Skew
Inequality in the amount of data available per client, creating an imbalance of influence during aggregation and potentially biasing towards well-supplied clients.
Federated Multi-Task Learning
Paradigm treating each client as a distinct task with shared and specific parameters, allowing effective modeling of non-IID heterogeneity.
Clustered Federated Learning
Approach grouping clients into clusters based on the similarity of their data distributions, enabling learning more adapted to the non-IID context.
Meta-Learning in Federated Settings
Application of meta-learning techniques to enable rapid adaptation of models to each client's specific distributions in a non-IID federated environment.
Federated Distillation
Technique where clients share their predictions or logits rather than their model weights, reducing the impact of non-IID data heterogeneity.
Adaptive Aggregation
Dynamic aggregation mechanism adjusting client contribution weights based on the similarity of their local updates, mitigating the effects of non-IID data.
Proxy Dataset
Reference dataset maintained by the server to evaluate and calibrate local models in a non-IID context, improving aggregation quality.
Federated Domain Adaptation
Domain adaptation technique in a federated framework to align feature distributions between clients with non-IID data.