Data augmentation using continuous conditional generative adversarial networks for regression and its application to...

Area of Science:

• Spectroscopy
• Machine Learning
• Data Science

Background:

• Machine learning in spectroscopy is limited by insufficient spectral samples.
• Effective algorithms for simulating synthetic spectra from limited real spectra are lacking for continuous regression models.

Purpose of the Study:

• To introduce a continuous conditional generative adversarial network (CcGAN) for autonomous synthetic spectra generation.
• To address the challenge of limited spectral data in machine learning-assisted spectroscopy analysis.

Main Methods:

• Developed a continuous conditional generative adversarial network (CcGAN) to generate synthetic spectra.
• Utilized a small dataset from a self-interference microring resonator (SIMRR)-based sensor.
• Evaluated generated spectra using principal component analysis (PCA) and incorporated them into deep neural network (DNN) regression models for data augmentation.

Main Results:

• CcGAN successfully generated high-quality synthetic spectra from limited real spectral data.
• Principal component analysis could not distinguish between real and CcGAN-generated synthetic spectra.
• Incorporating synthetic spectra significantly enhanced the predictive performance of the DNN regression model.

Conclusions:

• CcGAN is a promising approach for generating high-quality synthetic spectra.
• The developed method offers a superior data augmentation effect for regression tasks in spectroscopy.
• This technique effectively overcomes the limitation of insufficient spectral samples in machine learning applications.