Classifier Training for Open Vehicle Door Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in training machine learning models to accurately predict the state of objects, such as open doors, due to the scarcity of labeled data, leading to difficulties in identifying positive examples and incorporating noisy samples in the training process.
Innovation Solution
A method is developed to train machine learning classifiers using log data from autonomous vehicles, where initial positive examples are identified and additional positive examples are automatically generated based on temporal proximity, allowing for the creation of high-quality training data without human labelers, enabling accurate classification of objects in specific states like open doors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning training methods are used with limited labeled data, then training time and computational resources are reduced, but classification accuracy deteriorates due to scarcity of positive examples
Solution Approach 1:
The system performs preliminary action by using semi-supervised learning to pre-train classifiers on unlabeled data before final supervised training. This preliminary phase creates initial predictions that are then used to generate synthetic labeled training examples, effectively preparing the model in advance with expanded training data before the final accuracy-critical training phase.
Solution Approach 2:
The system creates copies of training examples by generating synthetic labeled data from initial model predictions. Positive examples are copied and augmented by creating variations with different transformations (rotation, scaling, translation) and by generating additional samples from the same underlying patterns, thereby multiplying the effective training data without requiring additional manual labeling.
2Measurement precision
If more labeled training data is collected to improve classification accuracy, then positive examples become more abundant, but the cost and time for human labeling increase
Solution Approach 1:
The system implements self-service by enabling the training process to generate its own labeled training data. The semi-supervised learning approach allows the model to use its own predictions on unlabeled data to create synthetic labeled examples, which are then fed back into the training process. This self-bootstrapping mechanism eliminates the need for external human labelers to expand the training dataset.
Solution Approach 2:
The system performs preliminary action by using semi-supervised learning to pre-train classifiers on unlabeled data before final supervised training. This preliminary phase creates initial predictions that are then used to generate synthetic labeled training examples, effectively preparing the model in advance with expanded training data before the final accuracy-critical training phase.
3Quantity of substance
If unlabeled data is incorporated through semi-supervised learning, then effective training data increases, but the complexity of the training process increases
Solution Approach 1:
The training process is segmented into distinct phases: an initial semi-supervised pre-training phase followed by a final supervised fine-tuning phase. This segmentation allows the complex task of learning from unlabeled data to be separated from the precision-critical phase of final classification, managing complexity by breaking the process into manageable stages with different objectives.
Solution Approach 2:
The system performs preliminary action by using semi-supervised learning to pre-train classifiers on unlabeled data before final supervised training. This preliminary phase creates initial predictions that are then used to generate synthetic labeled training examples, effectively preparing the model in advance with expanded training data before the final accuracy-critical training phase.
4Measurement precision
If synthetic training examples are generated from initial predictions, then training data quality improves, but the risk of propagating initial model errors increases
Solution Approach 1:
The system applies partial action by using only the high-confidence predictions from the initial model to generate synthetic training examples. By setting a confidence threshold and only selecting predictions above this threshold, the system partially uses the initial predictions rather than all of them, thereby reducing the propagation of low-confidence errors while still benefiting from the high-confidence correct predictions.
Solution Approach 2:
The system performs preliminary action by using semi-supervised learning to pre-train classifiers on unlabeled data before final supervised training. This preliminary phase creates initial predictions that are then used to generate synthetic labeled training examples, effectively preparing the model in advance with expanded training data before the final accuracy-critical training phase.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a classifier to detect open vehicle doors. One of the methods includes obtaining a plurality of initial training examples, each initial training example comprising (i) a sensor sample from a collection of sensor samples and (ii) data classifying the sensor sample as characterizing a vehicle that has an open door; generating a plurality of additional training examples, comprising, for each initial training example: identifying, from the collection of sensor samples, one or more additional sensor samples that were captured less than a threshold amount of time before the sensor sample in the initial training example was captured; and training the machine learning classifier on first training data that includes the initial training examples and the additional training examples to generate updated weights for the machine learning classifier.