Multi-Class Object Classification with Weighted Partial Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training machine-learned multi-class object classification models using partially labeled data leads to significant model quality degradation due to the assumption that unlabeled regions do not contain objects, requiring exhaustive and costly labeling of all classes in image data.
Innovation Solution
A computing system and method that utilize a loss function with weighted loss signals based on the inclusion of object classes within the labeled subset, adjusting model parameters to reduce the impact of unlabeled classes, allowing training with partially labeled data without degrading performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If exhaustive annotation of all classes is performed to train multi-class object classification models, then model quality is maintained, but training cost and time become prohibitively expensive
Solution Approach 1:
The patent applies partial action by training the model with only a subset of class labels (partially labeled data) rather than requiring complete annotation of all classes. The loss function is designed to handle missing labels by applying weights based on label inclusion, allowing the model to learn from available labeled data without needing exhaustive annotation of every class in the dataset.
2Productivity
If partially labeled training data is used to reduce annotation costs, then training efficiency improves, but model quality degrades significantly
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different classes in the loss function based on their labeling status. Each class receives a weight in the loss function determined by whether it is included in the labeled subset, allowing the model to focus learning efforts on well-labeled classes while still incorporating information from unlabeled classes without degrading overall model quality.
Solution Approach 2:
The patent changes the parameters of the loss function by introducing weights that are dynamically adjusted based on the inclusion of object classes in the labeled subset. This parameter modification allows the training process to adapt to partially labeled data, maintaining model quality by appropriately weighting the contribution of each class to the overall loss.
3Device complexity
If unlabeled regions are assumed to not contain objects to simplify training, then training complexity is reduced, but classification accuracy deteriorates
Solution Approach 1:
The patent introduces an intermediary mechanism in the form of a modified loss function that mediates between labeled and unlabeled regions. Rather than making binary assumptions about unlabeled regions, the loss function gracefully handles missing labels by applying appropriate weights, allowing the model to learn from both labeled and unlabeled regions without requiring complex additional processing or assumptions.
Data Source
AI summary
Systems and methods of the present disclosure are directed to a computer-implemented method for training a machine-learned multi-class object classification model with partially labeled training data. The method can include obtaining image data depicting objects and ground truth data comprising a subset of object class annotations respectively associated with a subset of object classes of a plurality of object classes. The method can include processing the image data with the machine-learned multi-class object classification model to obtain object classification data. The method can include evaluating a loss function that evaluates a multi-class classification loss and adjusting one or more parameters of the multi-class object classification model based on the loss function.