Deep CollabNet

Aiming to improve the learning of deep neural networks, this paper proposes the CollabNet network, which consists of a new method for inserting new hidden layers in Deep FeedForward neural networks, inspired by stacked autoencoders. The new way of insertion is considered collaborative and seeks training improvement over approaches based on stacked autoencoders. In this new approach, the layer insertion is performed in a coordinated and gradual way, keeping under the designer's control the influence of this new layer on training, and no longer in a random and stochastic way as in traditional stacking. The collaboration proposed in this work consists in making the learning of the newly inserted layer continue the learning obtained by the previous layers, without harming the network's global learning. In this way, the newly inserted layer collaborates with the previous layers and the ensemble works in a more aligned way with the learning. CollabNet was tested on a database of a real problem, obtaining satisfactory and promising results.