Neural Network Uncertainty Detection for Real-Time Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for autonomous driving systems struggle to accurately determine the uncertainty of predictions, especially in real-time applications where low latency is critical, and there is a lack of methods suitable for online use with limited computing power.
Innovation Solution
A method that modifies an artificial neural network to estimate uncertainty by determining an uncertainty range for each output data point, allowing for the separation of aleatoric and epistemic uncertainty, and using this information to adapt the system's responses and store data points for further training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ensembling methods or Bayesian neural networks are used to estimate predictive uncertainties, then uncertainty estimation accuracy is improved, but computational power requirements increase significantly
Solution Approach 1:
The patent segments uncertainty estimation into two distinct types: aleatoric uncertainty (data-dependent) and epistemic uncertainty (model-dependent). This segmentation allows each type to be handled with computationally efficient methods tailored to its specific characteristics, avoiding the need for computationally intensive ensembling or Bayesian methods while still providing accurate uncertainty estimates for both types.
Solution Approach 2:
The patent modifies the neural network output parameters to directly predict uncertainty values alongside predictions. By changing the output layer to include uncertainty parameters that can be computed directly from the network activations, the system achieves uncertainty estimation without requiring multiple network passes or complex Bayesian inference, thus reducing computational power consumption while maintaining accuracy.
2Reliability
If ensembling methods or Bayesian neural networks are used for uncertainty estimation, then reliability of uncertainty measurement is improved, but latency increases making online applications infeasible
Solution Approach 1:
The patent performs preliminary computation of uncertainty values during the forward pass through the neural network, so that both predictions and uncertainty estimates are available simultaneously. By computing aleatoric and epistemic uncertainty as byproducts of the standard forward propagation, the system eliminates the need for additional computational passes, thereby maintaining low latency suitable for online autonomous driving applications while preserving reliable uncertainty measurement.
3Productivity
If standard neural networks are used without uncertainty estimation, then computational efficiency is maintained, but the ability to detect corner cases and make conservative decisions is lost
Solution Approach 1:
The patent merges uncertainty estimation functionality directly into the standard neural network architecture by modifying the output layer to produce both predictions and uncertainty values in a single forward pass. This merging allows the system to maintain the computational efficiency of standard neural networks while simultaneously gaining the ability to detect corner cases through uncertainty thresholds, enabling conservative decision-making without sacrificing productivity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present invention relates to a method for detecting data points with high uncertainty in a machine learning (ML) model for an autonomous driving system (F). In an offline phase, an artificial neural network is provided (11), in which an uncertainty range is determined for each output data by the output layer. The modified neural network is trained with a training data set (12). Uncertainty measures for the data points of the training data set and, based thereon, uncertainty thresholds are determined (13). In an online phase, a new data point is acquired (14), and the uncertainty measure for the newly acquired data point is determined (15). A check is then performed (16) to determine whether the uncertainty measure determined for the newly acquired data point exceeds the uncertainty threshold determined in the offline phase.In the offline phase, the artificial neural network can be pre-trained for a task using a training data set (10), the pre-trained artificial neural network can be modified (11) by replacing the output layer of the artificial neural network with an output layer in which an uncertainty range is determined for each output data; and the modified artificial neural network can be re-trained (12) using the training data set that was already used for the original training before the modification of the artificial neural network.