Additional Funding Sources
This research is based upon work supported by the National Science Foundation Grant No. CHE-1904166, co-funded by CDS&E and the Office of Investigative and Forensic Sciences in the National Institute of Justice.
Abstract
The primary goal of spectral multivariate calibration is to determine a regression vector that relates spectral responses to their respective analyte values. Numerous regression techniques exist to fit a model to a collection of source samples with labeled analyte values, such as partial least squares (PLS). Models formed with traditional calibration can be used to predict unlabeled target samples, but a problem arises when target spectra and/or analyte amounts are shifted from the source domain in which the model was formed. Using transfer learning, models can be updated to target conditions and adapted to the target domain to predict more effectively. A variety of model updating methods for unlabeled target data exist but often possess several mathematical restrictions that limit their use. By leveraging constant analyte values of target samples and repeat spectra in model updating, the presented model updating method, null-augmentation regression constant analyte (NARCA), bypasses many of the constraints in other methods. Several NIR datasets have been examined with NARCA, and its performance is reported for difficult cow milk, cow feed, and mango data updating situations where the target sample domains are shifted from the source environment both in terms of spectra and analyte amounts.
Utilizing Constant Analyte Target Sample Sets in Transfer Learning
The primary goal of spectral multivariate calibration is to determine a regression vector that relates spectral responses to their respective analyte values. Numerous regression techniques exist to fit a model to a collection of source samples with labeled analyte values, such as partial least squares (PLS). Models formed with traditional calibration can be used to predict unlabeled target samples, but a problem arises when target spectra and/or analyte amounts are shifted from the source domain in which the model was formed. Using transfer learning, models can be updated to target conditions and adapted to the target domain to predict more effectively. A variety of model updating methods for unlabeled target data exist but often possess several mathematical restrictions that limit their use. By leveraging constant analyte values of target samples and repeat spectra in model updating, the presented model updating method, null-augmentation regression constant analyte (NARCA), bypasses many of the constraints in other methods. Several NIR datasets have been examined with NARCA, and its performance is reported for difficult cow milk, cow feed, and mango data updating situations where the target sample domains are shifted from the source environment both in terms of spectra and analyte amounts.