ANN Out-of-Distribution Detection via Binary Classifier Pool
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial neural networks (ANNs) are vulnerable to out-of-distribution (OOD) data, leading to unreliable and overconfident output results when encountering unknown classes or anomalies not present in the training dataset, which can result in incorrect labeling and potentially dangerous consequences.
Innovation Solution
A method that employs a pool of binary classifiers trained on groups of classes to detect OOD signal samples without affecting the original accuracy of the ANN, using a threshold-based approach to provide an indicator signal for out-of-distribution data, thereby improving reliability and reducing computational and memory footprint.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a standard ANN classification model is used, then classification accuracy for training classes is improved, but reliability against out-of-distribution data deteriorates
Solution Approach 1:
The classification task is segmented into two independent binary classification problems: one for detecting in-distribution samples and another for detecting out-of-distribution samples. This segmentation allows the system to maintain high accuracy for known classes while simultaneously detecting unknown classes, resolving the contradiction between classification accuracy and reliability against OOD data.
Solution Approach 2:
A decision boundary indicator is introduced as an intermediary mechanism that evaluates whether input samples lie within or outside the learned decision boundaries. This intermediary provides a reliability metric that complements the classification output, enabling the system to maintain accuracy for training classes while identifying OOD data points that would otherwise cause unreliable predictions.
2Reliability
If complex OOD detection methods like input perturbation or contrastive learning are applied, then reliability against OOD data is improved, but device complexity and computational burden increase
Solution Approach 1:
The patent employs lightweight binary classification models with significantly fewer parameters than the main classification network. These simplified models are trained once during the offline phase and then deployed for rapid OOD detection during online operation, providing reliable OOD detection with minimal computational overhead and avoiding the need for complex real-time processing.
Solution Approach 2:
The binary classification models and decision boundary indicators are trained in advance during an offline training phase using representative data from both in-distribution and out-of-distribution classes. This preliminary action prepares the OOD detection mechanism beforehand, eliminating the need for computationally intensive operations during real-time classification and reducing online computational complexity.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A processing device (14) comprising memory circuitry (18) having stored therein a set of weight values (W) and a threshold value (T) and instructions which, when executed in the processing device (14), cause the processing device (14) to: apply a first artificial neural network, ANN processing (20) to a set of sensing signals (SD), producing as a result a set of compressed representations of the sensing signals (EMB). The first ANN processing (20) is trained to produce the set of compressed representations (EMB) using a set of training signals distributed according to a set of training classes having an integer number L of classes. The instructions further cause the processing device (14) to: configure (BC) weight values of a plurality of computing units of a set of ANN processing circuits as a function of a set of weight values (W); apply a further artificial neural network, ANN processing (50) to the set of sensed signals (SD) via the set of ANN processing circuits configured as a function of the set of weight values (W), obtaining as a result a set of probability values indicative of a probability that a respective compressed representation in the set of compressed representations of sensing signals (EMB) corresponds to sensing signals in the set of sensing signals (SD) that fail to classify as belonging to the set of training classes having the integer number L of classes; obtaining an aggregated probability value (54) and providing (56) an indicator signal (D) based on a comparison (56) of the aggregated probability value and a threshold value (T).