Adaptive Data Augmentation for Deep Learning Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Acoustic event and scene recognition models trained under limited conditions perform poorly in real environments due to insufficient data diversity, leading to inconsistent recognition performance.
Innovation Solution
A method for training deep learning models that applies adaptive data augmentation based on recognition performance evaluation metrics, enhancing the model's ability to generalize by determining whether to augment training data based on overall and class-specific recognition performance thresholds, and applying data augmentation schemes such as mix-up, random cropping, and SpecAugment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If acoustic data is collected under limited conditions for training, then data collection is efficient and simple, but recognition performance deteriorates in real environments with various conditions
Solution Approach 1:
The patent applies data augmentation techniques during the training phase to pre-process and expand the training dataset. By generating synthetic training samples through transformations (noise addition, time shifting, pitch modification) before model training, the system prepares the model in advance to handle diverse real-world conditions, thus resolving the contradiction between limited data collection and robust real-environment performance
Solution Approach 2:
The patent modifies parameters of existing acoustic data to create varied training samples. By changing parameters such as adding different noise levels, adjusting time offsets, modifying pitch, and applying various signal processing transformations, the system expands the effective training data coverage without collecting new data under multiple conditions, thereby maintaining collection efficiency while improving real-environment reliability
2Reliability
If data augmentation is applied to enhance model generalization, then recognition performance in real environments improves, but training complexity and computation time increase
Solution Approach 1:
The patent implements selective data augmentation by applying different augmentation strategies to different classes based on their specific performance needs. Rather than uniformly augmenting all data, the system identifies classes with poor recognition performance and applies targeted augmentation to those specific classes, reducing overall training complexity while maintaining effectiveness for critical classes
Solution Approach 2:
The patent divides the data augmentation process into class-specific segments. By separating the training data into different classes and applying tailored augmentation techniques to each class based on its characteristics and performance requirements, the system manages training complexity through modular, organized processing rather than handling all data uniformly
3Ease of manufacture
If uniform data augmentation is applied to all classes, then implementation is simple, but performance deviation between different classes remains high
Solution Approach 1:
The patent applies different data augmentation strategies to different classes based on their specific characteristics and performance needs. Rather than using a uniform approach, the system tailors augmentation techniques to each class's requirements, applying stronger augmentation to classes with poor performance and milder augmentation to already-performing classes, thus achieving consistent cross-class performance while maintaining reasonable implementation complexity
4Reliability
If more training data is collected under various conditions, then model generalization improves, but data collection cost and time increase significantly
Solution Approach 1:
The patent creates synthetic copies of existing training data through data augmentation techniques. By generating multiple transformed versions of each original training sample (noise-added versions, time-shifted versions, pitch-modified versions), the system effectively multiplies the training dataset without requiring additional data collection, thus improving model generalization while avoiding the time cost of collecting diverse real-world data
Data Source
AI summary
Disclosed are methods of training a deep learning model and predicting a class and an electronic device for performing the methods. A method of training a deep learning model may include identifying training data labeled for each class, determining whether to augment the training data based on overall recognition performance indicating prediction accuracy of the deep learning model calculated in a previous epoch, augmenting the training data based on class-specific recognition performance indicating class-specific prediction accuracy of the deep learning model calculated in the previous epoch, predicting a class by inputting the training data or the training data that is augmented to the deep learning model according to a determination of whether to augment the training data, and training the deep learning model based on a labeled class and the predicted class.