Deep Neural Network Confidence via Standard Deviation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques used in vehicles for object classification and parameter estimation often lack access to ground truth data, leading to incorrect classifications and estimations in real-time, particularly when determining trailer angles.
Innovation Solution
The implementation of a system that utilizes multiple deep neural networks, including convolutional neural networks, to estimate trailer angles and calculate a standard deviation of predictions, which improves prediction accuracy by comparing the standard deviation to a predetermined threshold and enabling or disabling autonomous vehicle modes accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple deep neural networks are used to estimate trailer angles, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The system divides the single neural network into multiple separate deep neural networks (first DNN, second DNN, third DNN), each independently estimating trailer angle parameters. This segmentation allows the system to achieve more accurate measurements through multiple independent estimations while managing complexity by keeping each individual network relatively simple.
Solution Approach 2:
The system combines the outputs of multiple independent neural networks by calculating the standard deviation of their predictions. This merging approach integrates the results from multiple sources to produce a final measurement with associated confidence levels, improving overall measurement precision without requiring each individual network to be overly complex.
2Reliability
If standard deviation calculation is implemented to provide confidence levels, then reliability is improved, but loss of time increases due to additional processing
Solution Approach 1:
The system calculates only the standard deviation of the neural network outputs rather than performing complete re-evaluation or additional verification steps. This partial action approach provides sufficient confidence assessment for reliable operation while minimizing additional processing time, avoiding excessive computation beyond what is necessary for confidence evaluation.
3Object-affected harmful factors
If autonomous vehicle modes are disabled based on high standard deviation, then safety is improved, but productivity decreases due to mode restrictions
Solution Approach 1:
The system dynamically adjusts vehicle operation modes based on real-time confidence levels derived from standard deviation calculations. When confidence is high (low standard deviation), autonomous modes remain enabled for normal operation. When confidence is low (high standard deviation), the system transitions to safer modes or disables autonomous functionality. This dynamic adaptation ensures safety while maximizing productivity during reliable operation.
Data Source
AI summary
A system comprises a computer including a processor and a memory, and the memory including instructions such that the processor is programmed to calculate a standard deviation of a plurality of predictions, wherein each prediction of the plurality of predictions is generated by a different deep neural network using sensor data; and determine at least one of a measurement corresponding to an object based on the standard deviation.