Vehicle Expert Neural Networks via Local Student-Teacher Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks for vehicle applications, such as driver assistance and automated driving, are limited by their size and computational requirements, necessitating high computational effort and hardware constraints, and require extensive labeled training data, which is costly.
Innovation Solution
A method and system utilizing a student-teacher learning approach, where a global neural network is trained on a backend server and specialized expert neural networks are trained locally on reconnaissance vehicles using sensor data, reducing the need for extensive labeled data and optimizing the neural networks for specific contexts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a neural network is used for autonomous driving decisions, then the decision-making capability is improved, but the computing resources and processing time required increase
Solution Approach 1:
The patent segments the original neural network into multiple specialized sub-networks, each trained to handle specific driving scenarios (e.g., lane changing, merging, intersection navigation). This segmentation reduces the computational burden on any single network while maintaining overall decision-making capability through coordinated operation of multiple specialized networks.
Solution Approach 2:
The patent applies compression techniques that modify the parameters of the neural network, including weight pruning (removing unnecessary connections), quantization (reducing precision of weights and activations), and knowledge distillation (transferring knowledge from a large network to a smaller one). These parameter changes reduce the network size and computational requirements while preserving essential decision-making functions.
2Measurement precision
If neural network models are updated frequently to improve performance, then the decision accuracy is improved, but the download time and data usage increase
Solution Approach 1:
The patent extracts and removes redundant or less important parameters from the neural network model during compression. By identifying and eliminating unnecessary weights and connections, the system reduces the model size that needs to be downloaded and stored, while retaining the essential parameters needed for accurate decision-making.
Solution Approach 2:
The patent uses knowledge distillation where a large, accurate teacher network is used to train a smaller student network. The student network learns to replicate the teacher's decision-making behavior with fewer parameters, creating a compressed copy that requires less download time and storage while maintaining similar accuracy.
3Volume of stationary object
If the neural network is compressed to reduce size, then the storage requirements are reduced, but the processing precision may deteriorate
Solution Approach 1:
The patent systematically modifies network parameters through controlled pruning and quantization. By carefully selecting which parameters to remove and how to quantize remaining parameters, the system achieves compression while monitoring and maintaining processing precision through validation on test datasets.
Solution Approach 2:
The patent implements feedback mechanisms where the compressed network's performance is continuously evaluated against the original network's performance. If precision degradation is detected, the compression process is adjusted by retaining additional parameters or using less aggressive compression techniques, ensuring that storage reduction does not come at the cost of unacceptable precision loss.
Data Source
Figure 1
Figure 2
AI summary
The invention relates to a method for providing at least one compressed and specialized neural network (6) for a vehicle (50), wherein a global neural network (5) is trained by means of a training dataset (7) on a backend server (2), wherein the trained global neural network (5) is transmitted to at least one reconnaissance vehicle (60) of a vehicle fleet (70), wherein at least one expert neural network (6) is trained by means of sensor data (52), which were each captured in the at least one reconnaissance vehicle (60) by means of at least one sensor (51) in a context (30-x) corresponding to a specialization of the expert neural network (6), by means of a student-teacher learning method locally on the at least one reconnaissance vehicle (60), wherein an inference result of the global neural network (5) is used on the captured sensor data (52) in this instance as a respective basic truth, and wherein the at least one trained expert neural network (6) is provided. Furthermore, the invention relates to a system (1) for providing at least one compressed and specialized neural network (6) for a vehicle (50).