Sample Grouping for AI Model Adaptation With Hard-Easy Mining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The construction of AI models for medical images is costly, time-consuming, and labor-intensive due to the extensive resource expenditure on collecting and labeling data, necessitating a more efficient method for model building.
Innovation Solution
A mining method for sample grouping that involves obtaining a field dataset, removing outlier samples, and categorizing the remaining samples into hard and easy sets based on estimated and actual labeled results, using statistical distributions and normalization to optimize model building with a small sample number.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If extensive resources are spent on collecting and labeling medical images, then the AI model can be built with sufficient data, but the process becomes costly and time-consuming
Solution Approach 1:
The patent segments the sample dataset into three distinct groups: easy samples, hard samples, and outlier samples. This segmentation allows selective use of samples based on their difficulty and quality, enabling the model to be trained efficiently without requiring all samples to be labeled manually, thus reducing time and resource expenditure while maintaining model accuracy.
Solution Approach 2:
The patent performs preliminary actions by using an existing model to pre-label samples and calculate their difficulty scores before the actual model training begins. This pre-processing step identifies which samples are easy to label and which are hard, allowing the training process to focus resources strategically, thereby reducing overall labeling time and cost.
2Ease of manufacture
If existing models are used for medical image interpretation, then model building is simplified, but the accuracy may be degraded due to domain differences
Solution Approach 1:
The patent changes the parameter of sample selection by introducing a difficulty score threshold mechanism. Samples are filtered and selected based on their difficulty scores and similarity to the existing model's classification probability characteristics. This parameter-based approach allows the model to adapt to the specific domain while maintaining the simplicity of using pre-existing model architectures.
Solution Approach 2:
The patent implements feedback by using the existing model's output (estimated results) to evaluate sample quality and determine which samples should be included in the training set. The difference between estimated and actual labeled results provides feedback that guides the selection of hard and easy samples, ensuring the model adapts to the specific domain while maintaining accuracy.
3Adaptability or versatility
If all samples are used for model training, then the model can capture diverse patterns, but the training time and computational resources increase
Solution Approach 1:
The patent applies local quality by treating different samples differently based on their characteristics. Easy samples with high similarity to the existing model's classification probability characteristics are prioritized, while hard samples and outliers are handled separately or excluded. This localized approach to sample treatment maintains the model's ability to capture diverse patterns while significantly improving training speed by focusing computational resources on the most informative samples.
Data Source
AI summary
A mining method for sample grouping is provided. The method includes the following steps. A field dataset including multiple samples is obtained, and each sample corresponds to an actual labeled result. The samples are respectively input to an existing model, so as to obtain the estimated results. An outlier sample set in the field dataset is removed based on a difference distribution of the estimated results and the actual labeled results, and the samples that remain in the field dataset form a remaining sample set. The remaining sample set is grouped into a hard sample set and an easy sample set based on the estimated results of the remaining sample set.