Neural Network Confidence Value Determination via Activation Signature Scaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks used in partially automated systems, such as self-driving vehicles, lack the ability to quantify their confidence in object recognition, making it difficult to ensure safety and reliability, especially when input data differs from the training dataset.
Innovation Solution
A method to determine a confidence value for object recognition by comparing the activation signature of the detected object with an averaged activation signature distribution from the training data, using a neural network's output images and scaling them to match the input image dimensions, allowing for the quantification of anomalies and reliability assessment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for object recognition in automated systems, then object detection capability is improved, but the ability to quantify confidence in detections deteriorates
Solution Approach 1:
The patent introduces an intermediary confidence assessment module that mediates between the neural network's detection output and the automated system's decision-making process. This module computes confidence values by comparing detected object features against reference data from training sets, providing a quantitative reliability measure without modifying the core neural network detection capability.
Solution Approach 2:
The patent implements a feedback mechanism where confidence values are computed based on comparisons between detected object characteristics and reference characteristics from training data. This feedback loop provides continuous reliability assessment, allowing the system to adjust its operational responses based on the quantified confidence levels of detections.
2Measurement precision
If the neural network is trained on a specific data set, then detection accuracy for training data distribution is improved, but performance on anomalous or out-of-distribution data deteriorates
Solution Approach 1:
The patent performs preliminary actions by pre-computing and storing reference object characteristics and activation patterns during the training phase. These reference data are saved for later comparison during operational detection, enabling the system to assess confidence for novel or anomalous objects by comparing against the pre-established training data distribution without requiring retraining.
Solution Approach 2:
The patent employs lightweight, computationally inexpensive comparison operations (such as distance metrics or similarity measures) to assess whether new detections conform to training data distributions. These simple, disposable comparison operations enable rapid adaptability checks without the computational burden of retraining or complex adaptation mechanisms.
3Device complexity
If plausibility checks are limited to sensor comparison or expert knowledge analysis, then system complexity is reduced, but the ability to detect anomalies in neural network predictions deteriorates
Solution Approach 1:
The patent replaces complex mechanical or expert-knowledge-based plausibility checking systems with a computational approach using neural network activation patterns and feature comparisons. This substitution maintains relatively simple system architecture while significantly enhancing anomaly detection capability by leveraging the neural network's internal representation space.
Data Source
AI summary
A method is indicated for determining a confidence value of an object of a class detected in an input image with the aid of a trained neural network, including: producing an activation signature for the class of the detected object using a plurality of output images of a layer of the neural network, the input image being provided to the input of the neural network; scaling the activation signature to the dimension of the input image; comparing an object portion of the scaled activation signature with an activation signature distribution of all objects of the same class of a training data set of the neural network in order to determine the confidence value.