Anomaly Classifier Using Dual Likelihood Detectors for OOD Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning-based classifiers, such as neural networks, face challenges in detecting anomalies outside their training data distribution, particularly in safety-critical applications like autonomous vehicles, where misclassification can lead to severe consequences.
Innovation Solution
A computer-implemented method using an anomaly classifier that compares input signals against both a reference data distribution and a specific data distribution using generative models, specifically flow-based machine learning models, to determine likelihoods and classify inputs as anomalous or not, thereby enhancing detection precision and system safety.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a classifier is trained with data from a predefined distribution (e.g., day-time images), then it achieves good classification performance on training data, but it fails to correctly classify data from different distributions (e.g., night-time images)
Solution Approach 1:
The solution segments the classification task into two independent components: (1) a classifier trained on the specific data distribution for normal classification, and (2) an anomaly detector trained on a different distribution to identify out-of-distribution inputs. This segmentation allows each component to specialize, resolving the contradiction between high accuracy on training data and adaptability to new distributions.
Solution Approach 2:
The anomaly detector acts as an intermediary layer between the classifier and the final decision-making process. It receives inputs that the classifier cannot properly handle and flags them as anomalies, enabling the system to gracefully handle distribution shifts without compromising the classifier's specialized performance on its training distribution.
2Productivity
If a classifier is used for safety-critical applications without anomaly detection, then the system operates simply and quickly, but it risks severe consequences from misclassifying anomalous inputs
Solution Approach 1:
The anomaly detector is trained in advance on data from distributions different from the classifier's training data. During operation, it performs a preliminary check on incoming inputs before they reach the classifier, identifying potential anomalies upfront. This preliminary action prevents misclassification of anomalous inputs while maintaining fast operation for normal inputs.
Solution Approach 2:
The system prepares for potential failures by having the anomaly detector ready to catch out-of-distribution inputs before they can cause harm. This cushioning mechanism ensures that even if the classifier makes errors on anomalous inputs, the system has a protective layer that identifies these cases beforehand, preventing severe consequences in safety-critical applications.
3Measurement precision
If an anomaly detection system uses multiple detectors and comparisons, then it achieves higher detection precision, but it increases system complexity
Solution Approach 1:
The anomaly detection system is segmented into two independent detectors, each with a clear and simple function: the classifier detects patterns from its training distribution, while the anomaly detector detects deviations from expected patterns. This segmentation maintains individual component simplicity while achieving high overall detection precision through their coordinated operation.
Solution Approach 2:
The anomaly detector serves multiple functions: it identifies out-of-distribution inputs, flags potential anomalies, and works with any classifier trained on a different distribution. This multi-functionality justifies the added complexity by providing a universal solution that can be applied across different classification tasks and data types.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method for classifying, whether an input signal (x), which comprises image and/or audio data, is anomalous or not with respect to a second data distribution using an anomaly classifier (60), wherein the anomaly classifier (60) uses a reference detector (81) and a second detector (82), the method for classifying comprising the following steps: • Providing the input signal (x) to the anomaly classifier (60); • In the anomaly classifier (60), providing the input signal (x) to the reference detector (81) and the second detector (82); • Obtaining a reference value (Lref) from the reference detector (81) based on the input signal, wherein the reference value (Lref) characterizes the likelihood of the input signal (x) to belong to a reference data distribution; • Obtaining a second value (L2) from the second detector based on the input signal, wherein the second value (L2) characterizes the likelihood of the input signal (x) to belong to the second data distribution; Providing an output signal (y60), which characterizes a classification of the input signal (x) as anomalous or not based on a comparison of the reference value (Lref) and the second value (L2).