Reproduce “Variational Autoencoder for Calibration: A New Approach” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Travis Barrett, Amit Kumar Mishra, Joyce Mwangama
Researchers at Universität Cape Town and Aberystwyth University developed a Variational Autoencoder (VAE) architecture that directly uses its latent space as a sensor calibration output. This VAE simultaneously reconstructs uncalibrated sensor inputs and provides calibrated measurements, achieving competitive accuracy for pollutants like CO on a multi-sensor gas dataset.
Key insights
- The proposed VAE successfully functions as both an autoencoder, reconstructing original sensor inputs, and a direct calibrator, producing accurate outputs from its latent space.
- Constraining the latent space to a calibration output enables the model to implicitly learn the complex relationships and transfer functions between uncalibrated sensor inputs and ground truth.
- Preliminary tests showed that including the Kullback-Leibler Divergence term in the loss function reduced performance, leading to its exclusion for these specific models, suggesting that standard KLD regularization might not always be optimal for non-Gaussian sensor data distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.