OOD Detection via Noise Projection Matrix
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OOD data detection methods require significant computational resources and memory, particularly due to the need for covariance matrices in calculating Mahalanobis distances, making them inefficient for devices with limited resources like edge devices.
Innovation Solution
The proposed OOD data detection apparatus calculates a noise influence level by injecting noise into intermediate outputs of a trained model, using a projection matrix to evaluate the noise influence without requiring memory-intensive covariance matrices, thereby reducing computational and memory costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Mahalanobis distance method is used for OOD detection, then detection accuracy is improved, but memory usage and computational cost increase significantly
Solution Approach 1:
The patent extracts only the essential statistical properties (mean and covariance) needed for OOD detection from the intermediate layer outputs, storing them as pre-computed statistics rather than retaining all intermediate outputs. This extraction approach maintains detection accuracy while significantly reducing memory requirements by keeping only the necessary statistical parameters.
Solution Approach 2:
Instead of computing Mahalanobis distance directly from stored intermediate outputs (which requires high memory), the patent inverts the approach by pre-computing and storing only the mean and covariance matrices, then using these compact statistical representations to perform OOD detection through simplified distance calculations.
2Measurement precision
If Mahalanobis distance method is used for OOD detection, then detection accuracy is improved, but computational cost increases
Solution Approach 1:
The patent performs preliminary computation of the mean and covariance matrices during the training phase or offline preprocessing stage. By pre-computing these statistical parameters beforehand, the actual OOD detection process only requires simple distance calculations using the pre-stored statistics, significantly reducing the computational cost during runtime while maintaining accurate detection performance.
3Productivity
If deep learning model is applied for classification, then classification performance is improved, but ability to detect OOD data deteriorates
Solution Approach 1:
The patent makes the deep learning model serve multiple functions: it performs both its primary classification task and simultaneously provides OOD detection capability. By extracting statistical properties from intermediate layers that are useful for both classification and anomaly detection, the model achieves dual functionality without requiring separate dedicated systems, thus maintaining high classification performance while gaining OOD detection ability.
Data Source
AI summary
An OOD data detection apparatus includes: an obtainment unit that obtains monitoring target data; an intermediate output calculation unit that calculates an intermediate output by applying a trained model to the monitoring target data; a projected-component calculation unit that calculates a projected component of the intermediate output to a parameter constituting the trained model; and a discrimination unit that discriminates as to whether the monitoring target data is OOD data based on the projected component.