Machine Learning Model Training with Confidence-Based Relabeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained with symmetric loss functions on noisy label data may not accurately distinguish between optimal models trained on clean and noisy data, and stochastic gradient descent may fail to find the optimal model, especially when compared to cross-entropy loss in classification problems.
Innovation Solution
A processor-implemented method that determines prediction loss and confidence for class labels using a first machine learning model, trains a second model based on this confidence, and updates parameters using a loss function from which the symmetric loss function is excluded, allowing for relabeling of training inputs if confidence is below a threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a symmetric loss function is used to train a machine learning model on data with label noise, then the model can be trained on noisy data, but the model cannot accurately distinguish between optimal models trained on clean and noisy data
Solution Approach 1:
The patent divides the training process into two distinct phases: first training a teacher model on noisy data with symmetric loss function, then training a student model on clean data with cross-entropy loss function. This segmentation allows each model to serve different purposes - the teacher model adapts to noisy data while the student model achieves accurate classification by learning from the teacher's predictions on clean data.
Solution Approach 2:
The teacher model acts as an intermediary between the noisy training data and the student model. The teacher model's predictions serve as pseudo-labels that guide the student model's training, enabling the student to learn accurate classification without directly exposing it to the noisy labels. This intermediary approach resolves the contradiction by decoupling the adaptation to noisy data from the learning of optimal classification boundaries.
2Productivity
If stochastic gradient descent is used to process large amounts of data, then processing capability is improved, but the optimal model cannot be guaranteed to be found
Solution Approach 1:
The patent performs preliminary action by first training the teacher model on the noisy data to establish a baseline understanding of the data distribution. This preliminary training enables the subsequent student model training to focus on learning accurate classification from the teacher's predictions, rather than simultaneously dealing with both noisy label adaptation and optimization, thus improving the reliability of finding an optimal model while maintaining scalability.
3Adaptability or versatility
If a symmetric loss function is used for classification, then the model can handle noisy labels, but it is more difficult than cross-entropy loss to find an optimal model
Solution Approach 1:
The patent segments the loss function application by using symmetric loss function only for the teacher model training phase, and cross-entropy loss function for the student model training phase. This segmentation allows the system to leverage the noise-handling capability of symmetric loss where needed while using the optimization-friendly cross-entropy loss for the final classification model, thereby reducing the overall complexity of finding an optimal model.
Data Source
AI summary
A processor-implemented method includes: determining a prediction loss based on class prediction data obtained by applying a first machine learning model to a training input and a class label with which the training input is labeled; determining a confidence of the class label based on the determined prediction loss; and training a second machine learning model using the training input based on the determined confidence.