In earthquake engineering, acceleration time histories are the fundamental input for the nonlinear analysis of a structure. However, an acceleration time history is complex, and thus, obtaining complete information about a ground motion is a frequently faced challenge among engineers. One of the solutions to this challenge is to study the characteristics of ground motions and obtain predictive relationships such that a ground motion can be obtained for the required scenario. In the recent past, various Ground Motion Models (GMMs) varying from simple to complicated functional forms have been developed to generate ground motion spectra using source, path and site characteristics. Thus, the main objective of generating a GMM is to use a minimal set of variables that can effectively represent ground motions. This implies that the ground motion data can be reduced from superficial to inherent dimensions with as little information as possible. In this regard, Principal Component Analysis (PCA) is considered as the conventional technique for the dimensionality reduction of the data. In this method, the data is represented closely by several uncorrelated linear combinations of variables called principal components (PCs) in place of a large set of variables. However, the relations between these PCs and the original dataset are assumed to be linear in PCA. Thus, nonlinear principal component analysis (NLPCA) transpired instead of PCA, as ground motions result from a nonlinear physical process. The present study focuses on the dimensionality reduction and feature extraction of ground motions using an artificial neural network model based on the NLPCA technique. In this study, the NLPCA approach is used to determine the intrinsic dimensionality of Fourier Amplitude Spectra (FAS) for three events: the 2008 Iwate Earthquake, the 1999 Chichi Earthquake and the 1992 Landers Earthquake. The results indicate that the intrinsic dimensionality of a 100-dimensional FAS of acceleration time histories is three, with the first three principal components (PCs) explaining at least 86% of the data, for all the three individual events considered in this study. The correlations of these PCs with parameters such as rupture distance and site class are studied. The observations made are further confirmed through a comparison of the PCs obtained for Response Spectra of the respective ground motions.