Improving the Training of Deep Convolutional Neural Networks for Art Classification: from Transfer Learning to Multi-Task Learning

Deep Convolutional Neural Networks have become the most popular algorithm for Computer
Vision (CV) problems. However, they are well known to be particularly hard to train. A large
set of possible hyperparameters, combined with the need for large amounts of training data,
can put serious constraints on their optimization procedure. In this paper, we explore several
strategies which can facilitate their training when classifying images representing heritage
objects, a CV area which is relatively unexplored.