Noise-Robust Feature Extraction via Stochastic Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models trained with ideal data may not perform well when applied to real-world data containing unnecessary information, such as noise or background images, as the desired signal is often buried in noise, and removing such information can lead to artifacts like blurring, and obtaining noise-free data is difficult.
Innovation Solution
A learning apparatus that extracts feature values from input data using multiple neural networks with different parameters, stochastically converts these features, calculates losses based on similarity and processing accuracy, and updates parameters to minimize losses, thereby generating a robust model capable of processing data with noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a trained model trained with ideal data is used for inference of processing target data including unnecessary information, then the model structure can be simple, but sufficient performance cannot be obtained
Solution Approach 1:
The patent applies preliminary action by pre-processing the input data to remove unnecessary information before feeding it to the trained model. The determination unit identifies and removes noise or background images from the processing target data, allowing the model trained on ideal data to perform accurate inference without requiring a more complex model structure.
2Manufacturing precision
If unnecessary information is removed in advance from processing target data, then data quality improves, but artifacts such as blurring occur
Solution Approach 1:
The patent uses feedback by implementing a determination unit that analyzes the processing target data to identify unnecessary information based on learned characteristics. This intelligent feedback mechanism distinguishes between actual signal components and noise, removing only the latter while preserving the former, thus avoiding artifacts like blurring that occur with simple removal methods.
3Reliability
If data that does not include unnecessary information is used for training, then model training is effective, but it may be difficult to obtain such data
Solution Approach 1:
The patent introduces an intermediary approach by training the determination unit to learn characteristics of unnecessary information from actual data containing noise. This intermediary training process enables the system to identify and remove noise without requiring access to clean, noise-free training data, thus solving the data acquisition problem while maintaining training effectiveness.
4Reliability
If multiple neural networks with different parameters are used to extract feature values, then robustness to noise improves, but processing complexity increases
Solution Approach 1:
The patent applies merging by combining the outputs of multiple neural networks that extract feature values with different characteristics. The synthesis unit integrates these diverse feature representations to create a comprehensive understanding of the input data, enhancing noise robustness while managing processing complexity through coordinated integration rather than independent processing.
Solution Approach 2:
The patent implements dynamics by adaptively adjusting the contribution of different neural network outputs based on the characteristics of the input data. The synthesis unit dynamically weights and combines feature values from multiple networks, allowing the system to optimize its processing complexity based on the actual noise levels and data characteristics encountered.
Data Source
AI summary
According to one embodiment, a learning apparatus includes processing circuitry. The processing circuitry generates a first converted feature values and a second converted feature values by stochastically converting at least one of first feature values and second feature values. The processing circuitry calculates a first loss related to similarity between the first converted feature values and the second converted feature values. The processing circuitry obtains a first processing result by processing based on one or more third parameters with respect to the first converted feature values. The processing circuitry updates a parameter of at least one of the first parameters and the third parameters such that a value based on the first loss and a second loss calculated from the first processing result and a label is minimized.