Self-Supervised Federated Learning for Automotive Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In autonomous vehicle applications, neural networks lack reliable supervision signals for training, leading to decreased accuracy of inferences when trained on local data, and existing Federated Learning methods do not effectively address this issue in safety-critical contexts like autonomous driving.
Innovation Solution
The system employs Federated Learning with self-supervision by collecting sensor data from vehicles, applying transformations to create training datasets, and training neural networks locally without human annotations, ensuring data privacy and improving inference accuracy through local model updates transmitted to servers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Federated Learning is applied to train neural networks in autonomous vehicles using local data, then user privacy is protected, but inference accuracy decreases due to lack of reliable supervision signals
Solution Approach 1:
The system implements self-supervised learning where the neural network generates its own supervision signals by predicting transformations of sensor data. The network applies transformations (rotations, translations, scaling) to input data and learns to predict these transformations, enabling self-training without human annotations or external supervision signals, thus maintaining privacy while improving accuracy
Solution Approach 2:
The system changes the training approach by introducing transformation parameters (rotation angles, translation vectors, scaling factors) as learnable targets. Instead of relying on labeled data, the network learns by optimizing its ability to predict these transformation parameters, fundamentally changing how supervision signals are obtained in the Federated Learning context
2Measurement precision
If large amounts of sensor data are collected for training, then inference accuracy can be improved, but communication overhead increases when transmitting training data to servers
Solution Approach 1:
The system extracts only the essential training components (model weights and gradient updates) from the abundant sensor data, transmitting only these compressed representations to the server instead of transmitting the raw sensor data itself. This extraction approach maintains training effectiveness while dramatically reducing communication overhead and energy consumption
Solution Approach 2:
The training process is segmented into local computation phases (where full sensor data is processed) and communication phases (where only model updates are transmitted). This segmentation allows the system to leverage large local datasets for accurate training while minimizing the frequency and volume of data transmission to servers
Data Source
AI summary
A method includes receiving, from one or more server computers through a communication network, an edge 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 includes applying a transformation to the identified first data item to generate a second data item to form a training dataset containing the first data item, the second data item, and a signal representing the transformation between the first data item and the second data item. The method further includes training with respect to the edge model on the training dataset and transmitting first data representing the trained edge model to the one or more server computers though the communication network.