Object Classification With Intermediate-Feature OOD Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OOD detectors for deep-learning models in autonomous vehicles are inefficient, impractical, and lack flexibility, particularly when dealing with anomalous or outlier objects not present in training data, leading to inaccurate and unreliable predictions that can compromise driving safety and efficiency.
Innovation Solution
Construct an OOD detector by modeling the density of intermediate features within the deep neural network using a Gaussian mixture model (GMM) during model training, and finetune the model with regularization towards the original weights of a pretrained foundation model to improve separation between in-distribution and out-of-distribution inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing OOD detectors are used for deep-learning models in autonomous vehicles, then object classification can be performed, but the detection efficiency and reliability deteriorate when dealing with anomalous or outlier objects not present in training data
Solution Approach 1:
The patent segments the feature space by training separate Gaussian Mixture Models for in-distribution and out-of-distribution features. This segmentation allows the system to reliably distinguish between familiar objects (handled by ID detector) and anomalous objects (handled by OOD detector), resolving the contradiction between reliability for standard objects and adaptability for anomalous objects.
Solution Approach 2:
The patent introduces intermediate features extracted from the deep neural network as a mediator between the input data and the classification output. These intermediate features are fed to both the ID detector and OOD detector, enabling the system to maintain reliability for standard objects while gaining adaptability for anomalous objects through the OOD detection mechanism.
2Measurement precision
If machine learning models are trained on empirical sensor data, then accurate predictions can be made for training data distributions, but accuracy deteriorates for out-of-distribution inputs encountered in real-world driving
Solution Approach 1:
The patent performs preliminary action by training a separate OOD detector on intermediate features from the training data before deployment. This pre-trained OOD detector is ready to identify anomalous objects in real-time without requiring retraining, thus maintaining prediction accuracy for training data while enabling generalization to unseen objects through OOD detection.
Solution Approach 2:
The patent changes the parameter perspective by shifting from directly classifying input data to analyzing the distribution of intermediate features. By modeling the density of intermediate features using GMMs and comparing likelihood ratios, the system maintains precision for known distributions while adapting to unseen distributions through statistical feature analysis.
3Reliability
If OOD detectors are added to improve anomaly detection, then reliability for anomalous inputs improves, but system complexity increases
Solution Approach 1:
The patent uses copying by creating a parallel OOD detection pathway that mirrors the main classification pathway. The same intermediate features are copied and fed to both the ID detector and OOD detector, allowing anomaly detection without duplicating the entire deep neural network architecture, thus improving reliability while controlling complexity.
Solution Approach 2:
The patent simplifies the OOD detector by changing parameters from complex deep learning models to simpler Gaussian Mixture Models that only need to model the distribution of intermediate features. This parameter simplification reduces computational complexity while maintaining detection reliability through statistical feature density comparison.
Data Source
AI summary
An out-of-distribution (OOD) detector may be provided to efficiently identify OOD inputs to trained machine learning models, to prevent passing misleading predictions from the trained models to downstream systems. During training of a machine learning model including a deep neural network (e.g., an object detection and/or object classifier model based on sensor data), an out-of-distribution detector may be constructed by modeling the density of the intermediate features of the deep neural network (e.g., using a Gaussian mixture model). In some examples, the intermediate features may be within a finetuned model based on a pretrained foundation model, and the machine learning model may be finetuned with regularization toward the original weights of the pretrained model. During inference, the machine learning model may output predictions, such as multilabel object detection and/or classification outputs, and the out-of-distribution detector may be used to efficiently generate confidence values associated with the outputs of the model.