Anomaly Classifier Using Dual Distribution Likelihood Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning-based classifiers, such as those used in autonomous vehicles, struggle to detect anomalies outside their training data distribution, leading to potential misclassification and safety issues when encountering data from unfamiliar distributions.
Innovation Solution
A computer-implemented method using a dual generative model approach, where an input signal is evaluated by both a reference detector and a second detector, each modeling different data distributions, to determine if the signal is anomalous based on likelihood comparisons, leveraging flow-based machine learning models for accurate representation of high-dimensional data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a classifier is trained with data from a predefined distribution, then it achieves good classification performance on training data, but it fails to correctly classify data from different distributions (anomalies)
Solution Approach 1:
The system segments the detection task into two independent detectors: a reference detector trained on reference data distribution and a second detector trained on second data distribution. Each detector specializes in one distribution, and their results are combined to achieve both high accuracy on training data and adaptability to detect anomalies from different distributions.
Solution Approach 2:
The system introduces an intermediary anomaly classifier that combines the outputs of the reference detector and second detector. This intermediary component mediates between the two specialized detectors and provides the final anomaly classification, enabling the system to handle multiple distributions effectively.
2Measurement precision
If a single detector is used for anomaly detection, then the system complexity is low, but the detection precision for distinguishing different data distributions is insufficient
Solution Approach 1:
The detection system is segmented into multiple specialized detectors (reference detector and second detector), each trained on specific data distributions. This segmentation enables more precise anomaly detection by comparing likelihoods across different distributions, despite increasing system complexity.
3Productivity
If a classifier is trained only on specific data (e.g., day-time images), then it achieves high performance on that specific data, but it cannot handle out-of-distribution data (e.g., night-time images)
Solution Approach 1:
The system performs preliminary training of both the reference detector and second detector on their respective data distributions before deployment. This preliminary action ensures that each detector is well-adapted to its training distribution, enabling high classification performance while the comparison mechanism provides out-of-distribution detection capability.
Data Source
AI summary
A computer-implemented method for classifying whether an input signal, which comprises image and/or audio data, is anomalous or not with respect to a second data distribution using an anomaly classifier. The method includes: providing the input signal to the anomaly classifier; in the anomaly classifier, providing the input signal to a reference detector and a second detector; obtaining a reference value from the reference detector based on the input signal, the reference value characterizing the likelihood of the input signal to belong to a reference data distribution; obtaining a second value from the second detector based on the input signal, the second value characterizing the likelihood of the input signal to belong to the second data distribution; and providing an output signal, which characterizes a classification of the input signal as anomalous or not based on a comparison of the reference value and the second value.