Adaptive Aggregation Weights for Federated Learning Convergence

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Federated learning systems face challenges in managing data flow, training models, and privacy issues, particularly with non-independently and identically distributed (non-IID) data across collaborators, leading to performance degradation and convergence issues due to heterogeneity in local data distributions.

Innovation Solution

An adaptive aggregation method that calculates model divergence values for each collaborator, using a preserved test dataset to adjust aggregation weights based on divergence, class imbalance, and data distribution, ensuring robust aggregation and improved learning performance across heterogeneous data sets.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If standard federated averaging is used to aggregate model parameters, then the aggregation process is simple and computationally efficient, but the model convergence deteriorates due to non-IID data distribution across collaborators

Engineering Contradiction:
Improveaggregation efficiencyVSAvoidmodel convergence
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent changes the aggregation parameters by introducing adaptive weights based on loss values and gradient norms. Instead of uniform averaging, each collaborator's model parameters are weighted according to their local data quality and model performance, resolving the contradiction between simple aggregation and reliable convergence

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements feedback mechanisms where the server calculates loss values and gradient norms from each collaborator's local training, then uses these metrics to adjust aggregation weights. This feedback loop enables the system to adapt to non-IID data distributions while maintaining aggregation efficiency

Inventive Principle:
Principle #23Feedback

2Device complexity

If equal weights are assigned to all collaborators during aggregation, then the aggregation process is straightforward, but the model accuracy deteriorates when collaborators have heterogeneous data distributions

Engineering Contradiction:
Improveaggregation complexityVSAvoidmodel accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies local quality by assigning different aggregation weights to different collaborators based on their local data characteristics. Collaborators with higher quality data (lower loss, more stable gradients) receive higher weights, improving model accuracy without significantly increasing system complexity

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent dynamically changes aggregation parameters (weights) based on local training metrics. This allows the system to maintain simple aggregation logic while achieving high accuracy through adaptive parameter adjustment

Inventive Principle:
Principle #35Parameter changes

3Loss of energy

If model parameters from all collaborators are aggregated uniformly, then the computational overhead is low, but the learning performance deteriorates due to data heterogeneity

Engineering Contradiction:
Improvecomputational overheadVSAvoidlearning performance
Core Design Contradiction:
Loss of energyVSReliability

Solution Approach 1:

The patent introduces adaptive weighting parameters based on loss values and gradient norms, allowing the system to maintain low computational overhead while significantly improving learning performance through selective emphasis on high-quality local models

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4160488A1Adaptive aggregation for federated learning
Publication Date: 2023.04.05 SIEMENS HEALTHINEERS AG
  • EP4160488A1 patent drawingFigure 1
  • EP4160488A1 patent drawingFigure 2
  • EP4160488A1 patent drawingFigure 3

AI summary

Systems and Methods for adaptive aggregation in a federated learning model. An aggregation server (121) sends global model weights to all chosen collaborators (131) for initialization. Each collaborator (131) updates the model weights for certain epochs and then sends the updated model weights back to the aggregation server (121). The aggregation server (121) adaptively aggregates the updated model weights using at least a computed model divergence value and sends the aggregated model weight to collaborators (131).