Filter Pruning for Efficient Transfer Learning in Deep Convolutional Neural Networks

Convolutional Neural Networks are extensively used in computer vision applications. Many convolutional models became famous after being widely adopted in a variety of computer vision tasks because o their high accuracy and great generality. Trough Transfer Learning, pre-trained versions of these models can be applied to a large number of different tasks and datasets without the need to train an entire large convolutional model. We aim at finding methods to prune convolutional filters from these pre-trained models in order to make inference more efficient for the new task. To achieve this we propose a genetic algorithms based method for pruning convolutional filters of pre-trained models applied to a different dataset than the one they were trained for. After transferring knowledge from an already trained model to a new task, genetic algorithms are used to find good solutions to the filter pruning problem through natural selection. We then evaluate the results of the proposed methods and compare with state-of-the-art pruning strategies for convolutional neural networks. Obtained experimental results show that the method is able to maintain network accuracy while producing networks with a significant reduction in Floating Point Operations (FLOPs).