Transfer learning approaches for spectroscopic data
|Titel||Transfer learning approaches for spectroscopic data|
|Buchtitel||Conferentia Chemometrica, Collection of Abstracts|
A problem in the analysis of spectroscopic data gathered from chemical processes is the change of measurement conditions between calibration of a chemometric model and the time of application. These changes may be due to the change of the measurement setup, different environmental conditions or a modification of the measured substance itself (e.g. different recipe). In all cases, test data will be different compared to calibration data, which makes a direct application of the existing model impossible. Hence the goal is to get a model for the new situation that exploits as much of the information contained in the calibrated model or the data used to calibrate that model as possible. This idea is referred to as Transfer Learning in the machine learning literature . In chemometrics this problem is also known as calibration transfer. State of the art methods, which are e.g. implemented in the PLS toolbox, assume that the same set of samples is measured for both measurement set-ups . This may be a valid assumption for laboratory settings; for an in-line installation at production site, however, such a requirement can rarely – if at all – be fulfilled.