Neural Network Training via Quality-Based Sample Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The training of neural network models often requires large amounts of manually annotated data, leading to overfitting due to random sampling methods that reduce the data set size and proportion of high-quality samples, resulting in performance degradation.
Innovation Solution
A method that involves initial training using a first sample set, predicting on a second sample set, determining preferred samples based on prediction results, annotating and adding them to the first set, and iteratively updating the model until a training ending condition is met, thereby retaining valuable information and reducing annotation costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If random sampling learning is used to select training samples, then manual annotation costs are reduced, but data distribution is not changed and the proportion of high-quality samples remains the same, leading to overfitting and performance degradation
Solution Approach 1:
The patent changes the selection criterion from random uniform sampling to quality-based sampling. It introduces a quality assessment mechanism that evaluates each sample's information value, diversity, and representativeness, then selects samples based on these quality parameters rather than random probability, thereby improving model performance while maintaining reduced annotation costs
Solution Approach 2:
The system performs self-assessment of sample quality by automatically evaluating the information content and representativeness of each sample. This self-service mechanism allows the training process to identify and select the most valuable samples without external intervention, optimizing the training set composition based on intrinsic sample characteristics
2Reliability
If a large amount of training data is used, then model performance is improved, but manual annotation costs increase significantly
Solution Approach 1:
The patent extracts only the most valuable and representative samples from the entire dataset for manual annotation. By introducing a sample selection mechanism that identifies high-quality samples based on information content and diversity metrics, it extracts a small subset of critical samples that provide maximum training value, thereby reducing annotation time while maintaining model performance
Solution Approach 2:
Instead of annotating all available samples or using excessive random sampling, the patent applies partial action by annotating only the essential high-quality samples identified through quality assessment. This partial annotation approach is sufficient to train the model effectively without the need for complete dataset annotation
3Ease of manufacture
If the training sample set size is reduced through sampling, then annotation costs decrease, but the model is more likely to overfit
Solution Approach 1:
The patent changes the sampling strategy from uniform random selection to quality-based selection. It introduces multiple quality parameters including information content, diversity measure, and representativeness score. Samples are selected based on these parameters to maintain data distribution characteristics while reducing the total number of annotated samples, thereby preventing overfitting
Solution Approach 2:
The patent applies different quality criteria to different regions of the data space. It identifies and selects samples with high local quality characteristics (high information content, good diversity, strong representativeness) from various data distributions. This local quality approach ensures that each reduced sample set maintains the essential characteristics of the full dataset
Data Source
AI summary
The present disclosure relates to a method for training a neural network model performed at an electronic device. The method includes: performing initial training by using a first training sample set to obtain an initial neural network model; performing a prediction on a second training sample set by using the initial neural network model to obtain a prediction result of each of training samples in the second training sample set; determining a plurality of preferred samples from the second training sample set based on the prediction results; adding the plurality of preferred samples that are annotated to the first training sample set to obtain an expanded first training sample set; updating training of the initial neural network model by using the expanded first training sample set to obtain an updated neural network model until a training ending condition is satisfied.