Federated Learning Neural Architecture Search Optimization

The patent applies preliminary action by pre-calculating and storing performance scores for multiple candidate neural network architectures using synthetic data before the actual federated learning process. This pre-evaluation allows the system to quickly identify promising architectures without performing time-consuming training during the federated learning phase, thus resolving the contradiction between automatic optimization and search time.

The patent uses synthetic data as a copy or approximation of real data to pre-evaluate neural network architectures. By training and evaluating candidate architectures on synthetic data rather than real data, the system can perform multiple evaluations quickly to identify top candidates, significantly reducing the time required for architecture search while maintaining the automation benefit.

The patent performs preliminary evaluation of neural network architectures using synthetic data before the federated learning process begins. This pre-screening identifies a small subset of top-performing architectures that are then considered in the federated learning phase, reducing the number of architectures that need to be communicated and evaluated across distributed devices, thus lowering communication and aggregation overhead while maintaining evaluation accuracy.

The patent evaluates all candidate architectures on synthetic data (excessive action) to identify top performers, then only evaluates a limited subset of these top candidates on real local data (partial action). This approach ensures that the most promising architectures are identified with high accuracy while minimizing the communication and computational overhead during the actual federated learning process.