Hard Negative Mining via Score History for DNN Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) often struggle with class imbalance issues, where they tend to predict the dominant class even when unsure, due to low variation in training samples, leading to incorrect outcomes for certain samples.
Innovation Solution
The implementation of a method that generates an effective training epoch for DNNs based on score history from previous epochs, focusing on hard cases by maintaining scores for each sample and iteratively updating the training set to improve performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional training methods are used with standard training sets, then the training process is simple and fast, but the DNN produces incorrect outcomes for minority class samples due to class imbalance
Solution Approach 1:
The system performs preliminary identification of hard negative samples before final training by first training an initial DNN model, evaluating its predictions, and selecting misclassified samples as hard negatives. This preliminary action prepares a specialized training subset that addresses class imbalance issues before the main training process begins
Solution Approach 2:
The training set is segmented into different components: easy negatives, hard negatives, and positive samples. By dividing the training data into these segments and selectively emphasizing hard negatives in the training epoch, the system improves minority class classification without requiring complete retraining on all data
2Productivity
If the training set contains low variation in samples, then the training process is efficient, but the DNN predicts the dominant class even when unsure
Solution Approach 1:
The system implements feedback by evaluating DNN predictions on the training set, identifying samples where the model is uncertain or incorrect, and using these feedback results to select hard negative samples for the next training epoch. This feedback loop continuously improves the model's ability to handle uncertain cases
Solution Approach 2:
The training approach changes parameters by dynamically adjusting which samples are included in each training epoch based on their difficulty level. Hard negative samples that were initially excluded or underrepresented are selectively included in subsequent epochs, changing the composition parameters of the training set to improve prediction reliability
3Reliability
If all training samples are used in each epoch, then comprehensive coverage is achieved, but computational resources are wasted on easy samples that provide minimal learning value
Solution Approach 1:
Instead of using all training samples uniformly, the system applies partial action by selectively including only hard negative samples and relevant positives in each training epoch. This partial approach focuses computational effort on the most informative samples, reducing training time while maintaining or improving model performance
Solution Approach 2:
The training set composition is made dynamic by updating the selection of training samples in each epoch based on previous model performance. Samples that were hard negatives in one epoch may become easier in subsequent epochs, and the training set dynamically adapts to reflect the current model capabilities and learning needs
Data Source
AI summary
According to one embodiment, a method, computer system, and computer program product for hard negative training is provided. The embodiment may include a computer receiving a training set, where the training set comprises one or more training samples. The computer trains a deep neural network (DNN) with the training set. The embodiment may also include determining, using the DNN, information for each of the one or more training samples, where the information includes one or more scores associated with the one or more training samples. The embodiment may further include generating a training epoch from the one or more training samples based on the information and updates the information based on using the training epoch with the DNN.