Max-Margin, Single-Layer Adaptation of Transferred Image Features

Convolutional Neural Networks (CNNs) learned on the ImageNet dataset have been shown to be excellent feature extractors that, combined with linear SVM classifiers, yield outstanding results when transferred to target datasets (e.g., Pascal VOC, MIT Indoor 67 and Caltech) not used during the CNN learning process [?]. Given the large number of free parameters in CNN models (tens of millions), learning CNNs directly on these smaller target datasets is a difficult task. Yet recent work [?, ?] has established that it is possible to adapt the transferred CNN parameters to the smaller target dataset to further improve results.