Rare Example Mining in Driving Logs Using Embedding Density
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods fail to accurately identify and account for rare events or patterns in driving log data, particularly in unlabeled datasets, which are crucial for improving autonomous vehicle performance, as they do not consider varied properties within the same class label and rare events across time.
Innovation Solution
A system that uses an encoder neural network and density estimation model to compute a rareness measure from feature vectors in an embedding space, enabling the automatic mining of rare examples in large unlabeled data sets with minimal human input, thereby improving the performance of autonomous vehicles by incorporating these rare examples into training data for neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing methods are used to identify rare events in driving log data, then the system can process large datasets, but it fails to accurately identify rare events or patterns particularly in unlabeled datasets
Solution Approach 1:
The patent transforms the approach from direct rare event detection to density estimation in embedding space. By changing the parameter space from raw sensor data to learned feature embeddings, the system can identify rare events through density thresholds rather than explicit classification, enabling accurate detection in unlabeled data
Solution Approach 2:
The patent introduces an encoder neural network as an intermediary that maps sensor inputs to embedding space. This intermediary transformation allows the system to work with unlabeled data by learning meaningful representations first, then applying density-based rare event detection on the embedded features rather than raw data
2Reliability
If the system processes all driving log data to improve model performance, then rare scenarios can be captured, but computing resources required increase
Solution Approach 1:
The patent extracts only the rare examples from the large driving log dataset using density estimation. Instead of processing all data equally, the system identifies and extracts the minority dense regions (rare events) which are then used for targeted training, significantly reducing the computational burden while maintaining model performance on rare scenarios
Solution Approach 2:
The patent applies partial action by focusing computational effort only on identifying and processing rare examples rather than uniformly processing entire datasets. The density estimation approach allows selective extraction of rare events without exhaustive analysis of all data points, reducing overall computing resource requirements
3Device complexity
If existing methods consider only class labels for identifying rare events, then classification is simplified, but varied properties within the same class label are not considered
Solution Approach 1:
The patent moves the analysis from the single dimension of class labels to multiple dimensions by embedding sensor inputs into a high-dimensional feature space. The encoder neural network learns complex property variations within each class, and density estimation in this expanded space captures subtle differences that single-label classification would miss, without requiring explicit redefinition of classification categories
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing rare example mining in driving log data. In one aspect, a method includes obtaining a sensor input; processing the sensor input using an encoder neural network to generate one or more feature vectors for the sensor input; processing each of the one or more feature vectors using a density estimation model to generate a density score for the feature vector; and generating a rareness score for each of the one or more feature vectors from the density score. For example, the rareness score can represent a degree to which a classification of an object depicted in the sensor input is rare relative to other objects. As another example, the rareness score can represent a degree to which a predicted behavior of an agent depicted in the sensor input is rare relative to other objects.