Autoencoder With Unshared Hidden Units For Noise Cancellation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained with difference-source data from various acquisition devices and environments suffer from noise-related performance degradation due to inconsistent data quality, leading to high collection costs and reduced performance.
Innovation Solution
A data processing method using an autoencoder with unshared and shared hidden units to isolate and cancel noise, mapping data to a common feature space, and extracting common features for improved model robustness without the need for extensive data collection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If difference-source data from multiple acquisition devices and environments is used for training, then the data amount and coverage increase, but the noise interference increases and model performance degrades
Solution Approach 1:
The patent segments the feature extraction process into source-specific feature extractors and a shared classifier. Each source-specific extractor processes features from a particular acquisition device independently, allowing the system to handle multiple data sources while isolating source-specific noise characteristics. This segmentation enables the model to learn from diverse data without allowing noise from one source to interfere with the overall training process.
Solution Approach 2:
The patent extracts and removes source identification information from the feature vectors before inputting them to the classifier. By explicitly separating and removing the source identity component, the system prevents the classifier from learning source-specific noise patterns while retaining the useful information from difference-source data. This extraction approach directly addresses the noise interference problem while maintaining data diversity.
2Reliability
If more difference-source data is collected to improve robustness, then the training coverage increases, but the collection cost and difficulty increase
Solution Approach 1:
The patent creates a universal classifier that can process features from multiple different acquisition devices through a single unified model. The source-specific feature extractors serve as adapters that translate various source characteristics into a common feature representation that the universal classifier can process. This multi-functionality approach allows the system to achieve robustness across diverse data sources without requiring separate collection systems for each source type.
Solution Approach 2:
The patent uses source-specific feature extractors that can be trained on data from individual acquisition devices and then applied to process data from those same devices. Rather than collecting extensive data from every possible source, the system creates source-specific processing components that can handle the variability of each source type, effectively copying the adaptation process for each new data source while maintaining a shared classifier backbone.
3Ease of operation
If traditional training methods are used with difference-source data, then the training process is simple, but the model performance degrades due to noise
Solution Approach 1:
The patent segments the neural network into source-specific feature extractor modules and a shared classifier module. Each source-specific extractor is trained to process features from its corresponding acquisition device, while the shared classifier learns from the consolidated features from all sources. This segmentation maintains relative training simplicity through modular design while significantly improving model performance by preventing noise propagation from difference-source data.
Solution Approach 2:
The patent extracts source identification information from the input features and removes it before processing by the classifier. This extraction mechanism allows the system to maintain simple training procedures while improving performance by preventing the classifier from learning source-specific noise patterns. The source identification is extracted and handled separately, allowing the main training process to remain straightforward while achieving better generalization.
Data Source
AI summary
Disclosed are a data processing method, a device and a medium. The method includes: acquiring first feature data and a source identification of data to be processed; determining a first unshared hidden unit, corresponding to the source identification, in an autoencoder, wherein the autoencoder includes a plurality of first unshared hidden units that do not share a parameter with each other; inputting the first feature data into the determined first unshared hidden unit, to perform noise cancellation, and outputting second feature data meeting a set standard; inputting the second feature data into a first shared hidden unit of the autoencoder to map the second feature data to a set feature space through the first shared hidden unit, and outputting mapping data; and inputting the mapping data into a shared feature layer of the autoencoder, and outputting common feature data in the first feature data, extracted by the shared feature layer.