Teacher-Student Neural Network for Vehicle Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated driving systems face challenges in balancing computing power, memory, bandwidth, and accuracy with camera systems, where over-parametrization of deep learning neural networks consumes time and power, and the scarcity of data samples hinders training, leading to limited inference time.
Innovation Solution
A composed neural network architecture comprising a main neural network and a teacher neural network, where the teacher network is trained with a constant 'key' to provide weights to the main network, reducing over-parametrization and enabling efficient weight production only when the correct key is provided, thereby reducing inference time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If over-parametrization is used to train deep learning neural networks for object detection and classification, then the network can recognize various objects accurately, but it consumes excessive time and power during inference
Solution Approach 1:
The patent divides the neural network into two distinct components: a teacher network responsible for generating weights and a student network responsible for making predictions. This segmentation allows the weight generation process to be separate from the inference process, enabling the student network to operate with pre-computed weights and reduce inference time while maintaining detection accuracy through the teacher network's comprehensive training capability
Solution Approach 2:
The teacher network performs weight generation in advance before the actual detection task. By pre-computing and storing the optimal weights in a weight database, the system eliminates the need for real-time weight calculation during inference, thus reducing inference time and power consumption while preserving the accuracy benefits of over-parametrized training
2Reliability
If over-parametrization is used to train deep learning neural networks, then the network can learn from scarce data samples, but it requires large amounts of computing power and memory
Solution Approach 1:
The patent separates the computationally intensive weight generation task (performed by the teacher network) from the inference task (performed by the student network). This allows the system to use powerful computing resources during the offline weight generation phase while deploying a more resource-efficient student network for real-time operations, thus balancing training performance with power consumption constraints
Solution Approach 2:
The weight database acts as an intermediary between the teacher network and the student network. It stores the pre-computed weights that bridge the gap between the comprehensive training capability of the teacher network and the efficient inference capability of the student network, enabling the system to achieve high reliability with limited data while reducing the computing power requirements during deployment
3Power
If a small neural network is used to handle scarce data samples, then computing resources are conserved, but the network is hard to train and achieves lower accuracy
Solution Approach 1:
The teacher network serves as an intermediary that generates optimal weights for the smaller student network. This allows the student network to maintain its computational efficiency and resource efficiency while achieving high detection accuracy through the carefully crafted weights provided by the teacher network, effectively decoupling the trade-off between network size and accuracy
Solution Approach 2:
The patent changes the parameters of the student network by using pre-computed weights from the teacher network instead of training the student network from scratch with limited data. This parameter transfer approach enables the smaller network to achieve performance comparable to or exceeding that of larger networks while consuming fewer computing resources
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a system and a method for improving a detection algorithm of a driving assistance system. The system according to invention comprises at least one imaging device (10) for detecting objects in the surrounding of the vehicle, a processing device (20) and a controller device (30). The system is configured to: - receive image data from the imaging device/sensor (10), said image data regarding objects detected in the surrounding of the vehicle, - pre-process image data by means of a pre-processing device (210) and providing pre-processed image data, - process the pre-processed image data according to a trained model generated by means of a model training device (220) and providing detections, - post-process the detections by means of a post-processing device (230), - provide post-processed detections to the controller device (30), whereby a control decision is taken, - send the taken control decision to at least one processor of the vehicle driving system by means of communication means. The model training device (220) is part of the processing device (20) and has a composed neural network architecture that comprises: - a main neural network (f) designed to use weights (w) for mapping the pre-processed image to a detection output of the mentioned main neural network (f), and - a teacher neural network (fo) designed to use weights (w0) to produce weights (w) for the main neural network (f).