Filtering Server Training Samples for Federated Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning techniques face challenges such as catastrophic forgetting and difficulty in tuning parameters due to the need for separate optimization for each use case, particularly when combining server-based and client-based machine learning models.
Innovation Solution
A method that filters training data by comparing the losses of client and server machine learning models to select a subset of training samples, allowing the server model to be trained more efficiently without using gradients or tuning hyperparameters, and enabling quicker convergence and reduced computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If federated learning is used to fine-tune the global ML model using client gradients, then the model adapts to client data, but catastrophic forgetting occurs where information from server training is lost
Solution Approach 1:
The patent introduces a filtering mechanism as an intermediary between server training and federated fine-tuning. This filter selectively passes training samples from the server dataset to the federated learning process, preventing catastrophic forgetting by maintaining a balanced composition of training data that preserves both server-level and client-level information.
Solution Approach 2:
The patent dynamically adjusts the composition of training samples by changing parameters related to data selection. By modifying which samples are included in federated training based on loss comparisons, the system controls the balance between adapting to client data and retaining server training information, thus preventing catastrophic forgetting.
2Productivity
If traditional federated learning updates the global model using all client gradients, then the model converges, but hyperparameter tuning is required and computational resources are consumed
Solution Approach 1:
The filtering mechanism operates autonomously by comparing losses between server and client models to automatically determine which training samples should be used. This self-service approach eliminates the need for manual hyperparameter tuning and complex optimization procedures, simplifying the training process while maintaining efficiency.
Solution Approach 2:
The patent extracts only the necessary training samples from the full server dataset by applying a filtering criterion based on loss comparison. This extraction process removes unnecessary computational overhead and eliminates the need for processing all client gradients, thereby improving training efficiency without requiring extensive hyperparameter tuning.
3Reliability
If the server model is trained on the full server dataset, then the model achieves comprehensive knowledge, but computational resources and training time are excessive
Solution Approach 1:
The patent extracts a filtered subset of training samples from the complete server dataset by applying a loss-based filtering criterion. This extraction maintains the essential knowledge required for comprehensive model performance while significantly reducing the total number of training samples, thereby decreasing training time and computational resource requirements.
Solution Approach 2:
Instead of training on the full server dataset (excessive action), the patent applies partial action by selectively training on a filtered subset of samples that are most beneficial. This partial training approach achieves comparable or superior model performance with reduced computational cost and training time.
Data Source
AI summary
A method includes receiving, from a client device, a client machine learning (ML) model and obtaining a set of training data including a plurality of training samples. The client ML model is trained locally on the client device. For each respective training sample in the plurality of training samples, the method also includes determining, using the respective training sample, a first loss of the client ML model; determining, using the respective training sample, a second loss of a server machine learning (ML) model; and determining a respective score based on the first loss and the second loss. The method also includes selecting, based on each respective score of each respective training sample in the plurality of training samples, a subset of training samples from the plurality of training samples and training the server ML model using the subset of training samples.