Abstract
Initialization of neural network weights plays a pivotal role in determiningtheir performance. Feature Imitating Networks (FINs) offer a novel strategy byinitializing weights to approximate specific closed-form statistical features,setting a promising foundation for deep learning architectures. While theapplicability of FINs has been chiefly tested in biomedical domains, this studyextends its exploration into other time series datasets. Three differentexperiments are conducted in this study to test the applicability of imitatingTsallis entropy for performance enhancement: Bitcoin price prediction, speechemotion recognition, and chronic neck pain detection. For the Bitcoin priceprediction, models embedded with FINs reduced the root mean square error byaround 1000 compared to the baseline. In the speech emotion recognition task,the FIN-augmented model increased classification accuracy by over 3 percent.Lastly, in the CNP detection experiment, an improvement of about 7 percent wasobserved compared to established classifiers. These findings validate the broadutility and potency of FINs in diverse applications.