Abstract
Although recent approaches to combine machine learning (ML) enhanced models and adaptive model selections using the built models became popular in many areas, however, there could be crucial pitfalls of built models in terms of 1) feature extraction and importance, 2) unbalanced datasets and 3) different types of ML estimators. In this case study using the ImageNet dataset, we thoroughly investigate the effects and possible pitfalls of features, datasets, and ML estimators for adaptive model selection of CNN inference network models targeting embedded systems. Based on the comprehensive results and analysis, we summarize opportunistic alternatives to minimize the pitfalls in the model building phase using more accurate feature extraction algorithms, various datasets in balance and size, and different types of interpretable (explaining important features) and black-box models.