Encoding High-Level Features: An Approach To Robust Transfer Learning

Laurent Yves Emile Ramos Cheret, Thiago Eustaquio Alves de Oliveira

Abstract: Transfer Learning (TL) plays a vital role in image classification systems based on Deep Convolutional Neural Networks (DCNNs). Systems employing such technique may be susceptible to distortions on images, motivating the development of robust DCNNs capable of facing these problems. Unfortunately, changes in the architecture of DCNNs are sometimes specific to a kind of distortion and result in models that need to be retrained from scratch. This work proposes the use of autoencoders as intermediaries between pre-trained DCNNs and classifiers, delegating the denoising task to this architecture trained to encode feature maps. The classifiers are then trained to map the inputs from the autoencoder latent spaces to their respective classes. Models employing this approach achieved 3% to 4% increase in accuracy and 50% to 70% reduction in loss on the CIFAR10 and CIFAR100 datasets. The results also showed an up to 80% reduction in loss and up to 15% increase in accuracy for images with unseen distortions compared to the classical TL approach. This work improves classification results and increases robustness to distortions in a straightforward manner.