Federated Learning With Teacher Models for Vehicle Inference Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In autonomous vehicle applications, neural networks trained using Federated Learning lack reliable supervision signals, leading to decreased accuracy of inferences, especially in safety-critical contexts like autonomous driving, where large amounts of data are required for real-world accuracy and safety.
Innovation Solution
A system and method that incorporates a second neural network (teacher model) to derive inference signals as supervision signals for training, ensuring data privacy and increasing accuracy by using Federated Learning to aggregate updates from edge devices, with the second model being more robust and computationally heavier than the edge model for improved inference accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Federated Learning is used to train neural networks on distributed edge devices, then user data privacy is ensured, but the accuracy of inferences decreases due to lack of reliable supervision signals
Solution Approach 1:
A teacher model is introduced as an intermediary component that generates supervision signals locally on edge devices. This teacher model acts as a mediator between the available local data and the student model, providing the necessary supervision signals without requiring external human annotations or centralized data collection, thus maintaining privacy while improving training quality
Solution Approach 2:
The system enables self-service by allowing each edge device to generate its own supervision signals through the teacher model using locally available data. This eliminates the need for external annotation services or centralized training, allowing the system to improve inference accuracy independently at each edge device while maintaining data privacy
2Measurement precision
If large amounts of data are collected for training to ensure real-world accuracy and safety, then inference accuracy improves, but data transmission overhead and privacy risks increase
Solution Approach 1:
The supervision signal generation capability is extracted from the centralized server and embedded into the teacher model on each edge device. This allows each device to independently generate training supervision signals from its local data without needing to transmit the actual data or large amounts of training information to external servers, reducing transmission overhead while maintaining training quality
Solution Approach 2:
The training process is segmented into local operations at each edge device, where the teacher model generates supervision signals independently. This segmentation eliminates the need for centralized data aggregation and transmission, allowing each device to perform training locally with minimal communication overhead, thus reducing energy loss from data transmission
Data Source
AI summary
A method includes receiving, from one or more server computers through a communication network, a first model and collecting sensor data acquired by a sensor on a vehicle. The method also includes identifying a first data item from among the collected sensor data when the first data item is determined to satisfy a criterion. The method further include deriving an inference signal by running a trained second model using the first data item as input to the second model to provide a training dataset that contains the identified first data item and the derived inference signal as a supervision signal corresponding to the identified first data item. The method further includes training with respect to the first model on the training dataset and transmitting first data representing the trained first model to the one or more server computers though the communication network.