The Usefulness of Past Knowledge when Learning a New Task in Deep Neural Networks

In the current study we investigate the ability of a Deep Neural Network (DNN) to reuse, in a new task, features previously acquired in other tasks. The architecture we realized, when learning the new task, will not destroy its ability in solving the previous tasks. Such architecture was obtained by training a series of DNNs on different tasks and then merging them to form a larger DNN by adding new neurons. The resulting DNN was trained on a new task, with only the connections relative to the new neurons allowed to change. The architecture performed very well, requiring few new parameters and a smaller dataset in order to be trained efficiently and, on the new task, outperforming several DNNs trained from scratch.