Autonomous Vehicle Image Recognition With Perturbation Reliability Checks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information processing systems face high computational complexity and lack reliability evaluation when inputting data to trained models, particularly in neural networks with ReLU activations.
Innovation Solution
An information processing system that includes a first determination section to assess the change in output data relative to a pre-specified threshold after applying a perturbation to input data, and a control device to determine if new input data falls within a measured range, reducing computational complexity and evaluating output data reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If reachable set computation and safety verification for neural networks with ReLU activations is employed to precisely identify ranges of output data, then measurement precision is improved, but computational complexity increases significantly
Solution Approach 1:
The patent replaces the complex, computationally expensive reachable set computation method with a simpler, more efficient perturbation-based approach. Instead of using the original complex verification system, it employs a lightweight alternative that applies small perturbations to input data and measures output changes, achieving comparable precision with significantly reduced computational overhead.
Solution Approach 2:
The patent changes the fundamental parameters of the verification approach by switching from global reachable set computation to local perturbation-based analysis. It modifies the verification strategy by using epsilon-perturbed inputs and measuring output differences, fundamentally changing how the system approaches safety verification while maintaining precision requirements.
2Productivity
If the trained model is used to process input data in autonomous vehicles, then productivity is improved, but reliability of output data cannot be evaluated
Solution Approach 1:
The patent introduces a feedback mechanism where the system applies perturbations to input data, processes the perturbed data through the trained model, compares the output differences against a threshold, and uses this feedback to determine reliability. This closed-loop verification process enables reliability evaluation without significantly impacting the productivity of the main image recognition pipeline.
Solution Approach 2:
The patent performs preliminary reliability verification by applying perturbations and checking output stability before final output is generated. This preliminary action ensures that only reliable predictions are outputted, maintaining both productivity and reliability by filtering out uncertain results in advance.
Data Source
AI summary
An information processing device includes a first determination section and a first output section. First output data is obtained from a trained model, which has being trained beforehand, by input of first input data to the trained model. Second output data is obtained from the trained model by input of second input data to the trained model, in which second input data a perturbation of a specified perturbation amount is applied to the first input data. The first determination section makes a determination as to whether a change amount of the second output data relative to the first output data is not more than a pre-specified threshold. The first output section outputs information representing the first input data and a perturbation amount for which the change amount is determined by the first determination section to be not more than the threshold.