Will Transfer Learning Enhance ImageNet Classification Accuracy Using ImageNet-Pretrained Models?

The huge impact of Transfer Learning (TL) techniques in many fields was achieved using several state-of-the-art ImageNet-pretrained models. These models have shown great performance improvements on this dataset over the last few years. One of the recently used TL techniques is feature extraction with the help of Feature Concatenation (FC), where the extracted features of multiple pretrained models are concatenated together before training on them, to produce more robust and discriminative feature representations on various classification tasks. However, neither TL nor FC techniques have been tested on the same dataset that initially trained the pretrained models, i.e. ImageNet. Hence, this work provides an investigative study to test the possibility of improving the ImageNet accuracy using the feature extraction approach of TL with the help of FC techniques. The results of this work show that there is no TL technique that can be used with or without FC to increase the accuracy of pretrained models on the original dataset on which they were trained. Even for FC, it cannot produce a more discriminative feature representation for the original data than what the individual models can produce.