Development and Testing of Approach for Separated Training of Branches of Multi-Task Convolutional Neural Networks

This paper explores the possibility of building a new model on the basis of an already trained neural network to perform an extended number of tasks, using the knowledge available in the original model. Possible areas of application are determined, the influence of such an approach on the accuracy and speed of the obtained multi-task model is investigated in comparison with a set of single-task models each of which fulfills specific task.