Scheduled Federated Learning Device Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional federated learning methods face issues such as resource usage bias, dropout selection, and accuracy fairness due to random party selection, leading to inefficient resource utilization and biased model generation.
Innovation Solution
Implementing scheduled federated learning that pre-schedules devices based on availability and resource usage over time, using dynamic programming to optimize party selection and minimize resource usage variance, while prioritizing better contributing parties for improved accuracy and fairness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If random party selection is used in federated learning, then implementation simplicity is maintained, but resource usage bias and accuracy fairness deteriorate
Solution Approach 1:
The system performs preliminary actions by maintaining availability indicators and resource usage profiles for each computing device before the federated learning training rounds begin. This pre-characterization of devices allows the aggregator to make informed selections rather than random choices, resolving the contradiction by preparing advance information that enables both simplicity and fairness.
Solution Approach 2:
The invention changes the selection parameter from random selection to selection based on availability indicators and resource usage profiles. By transforming the selection criterion from a random parameter to a data-driven parameter, the system achieves both implementation simplicity (through automated scoring) and accuracy fairness (through balanced device selection).
2Device complexity
If random party selection is used in federated learning, then system complexity is minimized, but resource utilization efficiency deteriorates
Solution Approach 1:
Computing devices self-report their availability indicators and resource usage profiles to the aggregator. This self-service mechanism allows the system to gather necessary information without complex centralized monitoring, maintaining low system complexity while enabling efficient resource utilization through informed device selection.
Solution Approach 2:
The system implements feedback loops where device performance and resource usage are continuously monitored and fed back to the aggregator. This feedback enables dynamic adjustment of device selections across training rounds, improving resource utilization efficiency without significantly increasing system complexity through automated decision-making algorithms.
3Measurement precision
If more computing devices are selected for training rounds, then model accuracy is improved, but resource usage and dropout selection increase
Solution Approach 1:
The system applies partial action by selecting an optimal subset of computing devices for each training round rather than using all available devices. The availability indicators and resource usage profiles enable the aggregator to select just enough devices to achieve good model accuracy while avoiding the excessive resource consumption that would result from using all available devices.
Solution Approach 2:
The device selection is made dynamic through the use of availability indicators that can change over time. The aggregator can adapt the number and composition of selected devices based on current system conditions, resource availability, and training progress, allowing optimization of the balance between model accuracy and resource usage across different training rounds.
Data Source
AI summary
An indication of availability over time and resource usage is maintained for each computing device of a plurality of computing devices. An optimal combination of a subset of the plurality of computing devices is determined for each round of one or more rounds of training based on the availability over time and the resource usage for each computing device. A global model is generated utilizing the one or more optimal combinations of the plurality of computing devices and a query is performed utilizing the global model.