Mixed Client-Server Federated Learning for Catastrophic Forgetting

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Federated learning of machine learning models faces issues with catastrophic forgetting due to pre-training on biased data and challenges in obtaining diverse client gradients, leading to poor model performance.

Innovation Solution

A method where client devices generate client gradients based on local data, which are combined with remote gradients generated by a remote system to update global models, ensuring diverse and balanced data for more robust model updates.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If global ML models are pre-trained on proxy or biased data at the remote system, then the models can be initially trained without client gradients, but this results in catastrophic forgetting of information learned during pre-training

Engineering Contradiction:

Improveease of model trainingVSAvoidcatastrophic forgetting

Core Design Contradiction:

Ease of manufactureVSLoss of information

Solution Approach 1:

The system implements feedback loops where client gradients are continuously collected from deployed models and used to update the global model weights. This feedback mechanism ensures that the model adapts to actual client data distributions while retaining pre-trained knowledge through iterative refinement rather than complete retraining.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The system performs preliminary pre-training of global models on proxy data at the remote system before deployment. This preliminary action provides a good initialization that can be subsequently refined through federated learning, combining the benefits of efficient pre-training with adaptive fine-tuning on client data.

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

If weights of global ML models are updated only based on client gradients from deployed models, then the models adapt to client data, but this results in poor performance when certain client gradients (e.g., false positives, false negatives) are difficult to obtain

Engineering Contradiction:

Improvemodel adaptability to client dataVSAvoidmodel performance reliability

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The system merges multiple gradient sources including client gradients from deployed models, synthetic gradients generated by the remote system, and gradients from diverse client devices. This combination ensures comprehensive coverage of different data scenarios and improves model reliability by compensating for deficiencies in any single gradient source.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The system introduces an intermediary remote system that generates synthetic gradients and acts as a mediator between client devices and the global model. This intermediary provides additional gradient signals from diverse data sources, ensuring reliable model updates even when direct client gradients are unavailable or insufficient.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Productivity

If federated learning uses only client gradients from deployed models, then the learning process is communication-efficient, but this limits the diversity and balance of data used for model updates

Engineering Contradiction:

Improvecommunication efficiencyVSAvoiddata diversity for model updates

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The system segments the gradient generation process into multiple independent sources: client gradients from deployed models, synthetic gradients from the remote system, and gradients from diverse client devices. This segmentation allows each source to contribute specialized gradient signals while maintaining overall communication efficiency through selective aggregation.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system adds another dimension to gradient collection by incorporating synthetic gradients generated at the remote system alongside traditional client gradients. This dimensional expansion diversifies the data sources without requiring additional communication rounds, thereby improving data diversity while maintaining communication efficiency.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4244846B1Mixed client-server federated learning of machine learning model(s)

Publication Date: 2026.02.25 GOOGLE LLC

Figure 1A
Figure 1B
Figure 1C

AI summary

Implementations disclosed herein are directed to federated learning of machine learning ("ML") model(s) based on gradient(s) generated at corresponding client devices and a remote system. Processor(s) of the corresponding client devices can process client data generated locally at the corresponding client devices using corresponding on-device ML model(s) to generate corresponding predicted outputs, generate corresponding client gradients based on the corresponding predicted outputs, and transmit the corresponding client gradients to the remote system. Processor(s) of the remote system can process remote data obtained from remote database(s) using global ML model(s) to generate additional corresponding predicted outputs, generate corresponding remote gradients based on the additional corresponding predicted outputs. Further, the remote system can utilize the corresponding client gradients and the corresponding remote gradients to update the global ML model(s) or weights thereof. The updated global ML model(s) and/or the updated weights thereof can be transmitted back to the corresponding client devices.