Correcting Sensor-Induced Distribution Drift with Wasserstein Adversarial Learning
Researchers have developed a novel method using Wasserstein Generative Adversarial Networks (WGANs) to correct distribution drift in sensor data caused by factors like motion or aging. This approach treats the generator as a learnable calibration transformation, with the critic providing a distributional distance signal via the Wasserstein objective. The technique was successfully applied to simulated calorimeter data, accurately recovering aging coefficients and improving energy-sum distribution agreement, indicating its potential for data-driven calibration strategies where direct labels for degradation are absent. AI
IMPACT Introduces a new method for unsupervised calibration of sensor data, potentially improving the reliability of data-driven systems in physics and engineering.