Attribute-Based Calibration for Machine Learning Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning classification methods, such as Zero Shot Learning (ZSL) and Generalized Zero Shot Learning (GZSL), fail to accurately predict classes based on unseen attributes, as they rely on attributes from both seen and unseen classes during training, leading to domain shift issues and imbalanced datasets.
Innovation Solution
The introduction of an entropy-based penalty in the attribute-level loss function during training, which penalizes high-certainty predictions of zero probability for unseen attributes, enables the model to predict unseen attributes more accurately and mitigates overfitting by incorporating a weighted attributes-level entropy loss to address dataset imbalances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If ZSL or GZSL models are trained using attributes from both seen and unseen classes, then the model can predict unseen classes, but the model generates much higher scores for seen classes relative to unseen classes, causing domain shift and unreliable predictions for unseen classes
Solution Approach 1:
The patent applies parameter changes by modifying the loss function parameters during training. Specifically, it introduces an unseen class attribute loss component that adjusts the weighting and calculation parameters of the loss function to account for unseen classes, thereby balancing the prediction scores between seen and unseen classes and mitigating domain shift
Solution Approach 2:
The patent implements feedback by computing the unseen class attribute loss based on the difference between predicted and ground truth attributes for unseen classes, and using this feedback signal to update the model parameters during training, thereby improving the model's ability to predict unseen classes reliably
2Measurement precision
If the model is trained only on seen classes, then the model achieves high accuracy for seen classes, but the model cannot predict classes not seen during training
Solution Approach 1:
The patent applies segmentation by dividing the attribute prediction task into separate components: one for seen class attributes and another for unseen class attributes. This is achieved by computing distinct loss components (seen class attribute loss and unseen class attribute loss) that independently optimize predictions for each type of attribute
Solution Approach 2:
The patent implements universality by designing a unified loss function that simultaneously handles both seen and unseen class predictions. The total loss function combines multiple components (seen class attribute loss, unseen class attribute loss, and domain-invariant feature loss) to make the model universally applicable to both seen and unseen classes
3Productivity
If the model uses standard training approaches, then the training process is simple and fast, but the model fails to account for dataset imbalances and produces biased predictions
Solution Approach 1:
The patent applies parameter changes by introducing weighting parameters in the loss function to account for dataset imbalances. Specifically, it uses class frequencies and attribute importances as parameters to weight the loss components, thereby adjusting the training process to handle imbalanced datasets while maintaining computational efficiency
Data Source
AI summary
Machine learning classification using attribute-based calibration can include encoding a set of features extracted from computer-readable data associated with an object, the set of features describing one or more predetermined aspects of the object. A set of attribute predictions can be generated based on the set of features. The set of attribute predictions can be generated by a machine learning model that is capable of generating predictions for unseen attributes and that is trained using an attributes-level loss function. The attributes-level loss function can include an unseen attributes loss component that is computed only with respect unseen attributes. The set of attribute predications can be mapped to a set of predetermined attributes corresponding to one of a plurality of predetermined classes. An output of the machine learning classification is the classification of the object based on the mapping.