Attention-Map Labeling for Mixed Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data augmentation methods for machine learning, such as Mixup and CutMix, fail to accurately determine labels for mixed images due to unequal contributions of image pixels and background interference, leading to incorrect classifications and increased computational overhead.
Innovation Solution
The TransMix method generates labels for mixed images using an attention map based on the contributions of the original images, leveraging self-attention techniques in transformer models to dynamically weight the importance of image portions, thus improving label accuracy without additional parameters or significant computational cost.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is performed by mixing existing samples, then the quantity of training samples is increased, but the accuracy of labels for mixed data deteriorates
Solution Approach 1:
The patent introduces an attention map as an intermediary to bridge the mixed data and its label determination. The attention map quantifies the contribution of each original sample to the mixed data, serving as a mediator that enables accurate label assignment. By using this intermediary, the system can generate numerous mixed samples while maintaining label accuracy through attention-weighted label computation.
2Measurement precision
If manual methods are used to determine labels for mixed data, then label accuracy may be improved, but computational overhead and time consumption increase
Solution Approach 1:
The patent implements self-service by enabling the model to automatically determine its own attention maps and compute labels for mixed data without external manual intervention. The attention mechanism inherently provides the contribution weights needed for label computation, allowing the system to serve itself in the label determination process. This eliminates time-consuming manual labeling while maintaining accuracy through the mathematical relationship between attention weights and label probabilities.
3Measurement precision
If attention mechanisms are applied to determine labels for mixed data, then label accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent achieves universality by designing an attention mechanism that serves multiple functions simultaneously: it identifies important regions in the mixed data, quantifies sample contributions, and enables accurate label determination. This multi-functional approach avoids the need for separate computational modules for each task, thereby improving label accuracy while limiting the increase in overall computational complexity through efficient resource utilization.
Data Source
AI summary
Implementations of the present disclosure relate to methods, devices, and computer program products for data augmentation. In the method, mixed data is generated from first data and second data, and the mixed data comprises a first portion from the first data and a second portion from the second data. An attention map is obtained for the mixed data based on distributions of the first and second portions in the mixed data, here the attention map describes contributions of the first and second data to the mixed data. A label is determined for the mixed data based on the attention map and a first label for the first data and a second label for the second data. With these implementations, the label is determined based on the contributions of the first and second images in an accurate and effective way, and thus has a value that is much closer to the ground true.