Multiple Resolution DNNs for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous driving systems using single deep neural networks (DNNs) often suffer from accuracy and robustness issues, particularly in advanced driver assistance systems, due to the limitations of a single training dataset and potential noise in input data.
Innovation Solution
The system employs multiple resolution deep neural networks trained on sub-sampled datasets with different resolutions, using techniques like sequential leader clustering and varying dropout rates, and combines their outputs through a voting scheme to determine the best output, while also monitoring for degradation states to adjust or limit feature operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single deep neural network is used for autonomous driving features, then the system complexity is low, but the accuracy and robustness are insufficient
Solution Approach 1:
The training dataset is segmented into multiple sub-sampled datasets with different resolutions through sequential leader clustering. Each subset captures different characteristics of the data distribution, enabling multiple DNNs to learn complementary features and improve overall robustness without requiring a single overly complex model
Solution Approach 2:
Multiple DNNs are trained with different hyperparameters including varying dropout rates and learning rates. This parameter diversification allows each network to develop different inductive biases and generalization patterns, improving ensemble robustness while keeping individual network architectures relatively simple
2Reliability
If multiple deep neural networks are trained with different hyperparameters, then the robustness improves, but the training time and computational resources increase
Solution Approach 1:
The system trains multiple DNNs with varying degrees of regularization (dropout rates) rather than perfectly optimizing a single model. This partial action approach accepts that individual models may not achieve maximum performance, but the ensemble provides robustness with reasonable training time by avoiding excessive optimization of each component
Solution Approach 2:
Sequential leader clustering is performed beforehand to pre-process the training data into structured subsets before DNN training begins. This preliminary organization of data into resolution-based groups enables more efficient parallel training of multiple networks by reducing data preprocessing overhead during the training phase
3Reliability
If the training dataset is sub-sampled to create multiple resolutions, then the noise susceptibility is reduced, but the data quantity available for training decreases
Solution Approach 1:
The full training dataset is segmented into multiple sub-sampled subsets with different resolutions using sequential leader clustering. Each subset contains a representative portion of the data at a specific resolution level, allowing models to learn from varying levels of detail while collectively utilizing the entire original dataset across the ensemble
Solution Approach 2:
Instead of reducing data quantity, the system adds a resolution dimension by creating subsets at different granularity levels. This transforms the problem from having insufficient data to having rich multi-resolution data, where each resolution level provides complementary information that improves noise robustness
4Measurement precision
If multiple DNNs are executed simultaneously, then the output accuracy improves through voting, but the processing speed decreases
Solution Approach 1:
The ensemble of multiple DNNs is segmented into parallel processing units that can execute simultaneously. Each network processes inputs independently and contributes to the final voting result, enabling accurate ensemble predictions without sequential processing bottlenecks when sufficient computational resources are available
Data Source
AI summary
Techniques for training multiple resolution deep neural networks (DNNs) for vehicle autonomous driving comprise obtaining a training dataset for training a plurality of DNNs for an autonomous driving feature of the vehicle, sub-sampling the training dataset to obtain a plurality of training datasets comprising the training dataset and one or more sub-sampled datasets each having a different resolution than a remainder of the plurality of training datasets, training the plurality of DNNs using the plurality of training datasets, respectively, determining a plurality of outputs for the autonomous driving feature using the plurality of trained DNNs and the input data, receiving input data for the autonomous driving feature captured by a sensor device, and determining a best output for the autonomous driving feature using the plurality of outputs.