Multi-Class Model Repurposing for Multi-Label Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training multi-class classification models to accurately assign labels to unlabeled data objects is resource-intensive and costly, especially when large datasets are required, and existing models are not designed to handle non-mutually exclusive classifications, which limits their effectiveness in identifying data objects with shared characteristics across classes.
Innovation Solution
Reusing a trained multi-class classification model to generate non-mutually exclusive probability distributions for unlabeled data objects, allowing for the application of multiple labels by isolating instances of the model to determine binary confidence levels for each class, thereby reducing the need for new model training and resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a multi-class classification model is trained to accurately assign labels to unlabeled data objects, then labeling accuracy is improved, but resource consumption and training cost increase
Solution Approach 1:
The patent applies preliminary action by pre-training a multi-class classification model on labeled data before deploying it for labeling unlabeled data. The model is trained in advance using supervised learning with labeled training data, and then this pre-trained model is reused to generate probability distributions for unlabeled data objects, avoiding the need to retrain the model for each new labeling task and thus reducing resource consumption while maintaining accuracy
Solution Approach 2:
The patent uses copying by creating multiple independent instances of the same pre-trained classification model. Each instance independently processes different data objects or batches, generating probability distributions without requiring additional training. This allows parallel processing and efficient utilization of the trained model's knowledge across multiple labeling tasks, reducing both time and computational resources needed
2Productivity
If a multi-class classification model is used to assign labels, then classification speed is improved, but the model cannot handle non-mutually exclusive classifications, limiting effectiveness for data objects with shared characteristics
Solution Approach 1:
The patent applies segmentation by dividing the classification task into independent binary classification problems for each class. Instead of forcing a single-class assignment, the model generates separate probability estimates for each class independently. This allows data objects to be associated with multiple classes simultaneously based on their probability distributions, enabling handling of non-mutually exclusive classifications while maintaining fast inference speed
Solution Approach 2:
The patent changes the output parameter format from traditional single-class prediction to multi-class probability distributions. By outputting independent probability values for each class rather than a single class label, the model can represent complex class associations where data objects belong to multiple classes. This parameter change enables versatility in handling various classification scenarios including non-mutually exclusive cases
Data Source
AI summary
Techniques are disclosed relating to receiving, by a computer system, an unlabeled data object to be labeled using a classification model that is trained to output a probability distribution across a plurality of classes that are treated by the classification model as mutually exclusive. The technique may further include using, by the computer system, the classification model in a manner that determines a set of non-mutually-exclusive probabilities that respective ones of the plurality of classes apply to the unlabeled data object. Additionally, the technique may include labeling, by the computer system using the set of non-mutually-exclusively probabilities, the unlabeled data object.