Data Classification Model Disentangling Label Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based classification models are sensitive to the distribution of training data, leading to overfitting and decreased accuracy when the distribution of target data differs from that of the training data, especially in cases of non-uniform data distributions.
Innovation Solution
A method that disentangles the label distribution of source data from the classification model's training process, using a second equation to generate a first output value, and applies information about the target data's label distribution to classify data accurately, regardless of the data distribution, through techniques like Monte Carlo approximations and regularized Donsker-Varadhan representations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a classification model is trained using deep learning technology with source data, then the model can classify data according to the training distribution, but the accuracy decreases when the target data distribution differs from the source data distribution
Solution Approach 1:
The patent extracts the label distribution component from the training process by introducing a separate distribution parameter θ that is independently optimized. This allows the classification model to separate the learning of classification features from the learning of label distribution, enabling the model to adapt to different target distributions without retraining the entire model.
Solution Approach 2:
The patent changes the parameter representation by introducing a distribution parameter θ that explicitly models the label distribution. By optimizing θ separately and combining it with the base model output, the system can adapt to different data distributions through parameter adjustment rather than full retraining, thereby improving both accuracy and distribution adaptability.
2Reliability
If the classification model is trained to fit the source data distribution well, then the model achieves high accuracy on source data, but the model becomes sensitive to distribution shifts and performs poorly on target data with different distributions
Solution Approach 1:
The patent introduces a distribution parameter θ as an intermediary that mediates between the base classification model and the target data distribution. This intermediary allows the model to adapt to distribution shifts without being directly trained on target data, reducing sensitivity to distribution changes while maintaining reliable performance across different datasets.
Solution Approach 2:
The patent performs preliminary optimization of the distribution parameter θ using source data before applying it to target data. This preliminary action prepares the model for distribution shifts by pre-learning the distribution characteristics, enabling the model to maintain reliable performance when deployed on target data with different distributions.
Data Source
AI summary
A method of classifying data includes: training a classification model for classifying input data into at least one class, such that a first output value is generated according to a second equation in which a component corresponding to a label distribution of source data is disentangled in a first equation corresponding to the classification model; generating a second output value by applying, to the first output value, information indicating a label distribution of target data; and classifying the target data into the at least one class by using the second output value.