Inserção colaborativa de camadas em redes stacked autoencoder

Abstract

This work proposes an innovative method for the structured growth of deep neural networks of the stacked autoencoder type, aiming to overcome limitations of conventional network expansion approaches. Deep neural networks have proven highly effective in classification, pattern recognition, and automatic feature extraction tasks; however, defining appropriate topologies remains a challenge. In traditional methods, inserting new hidden layers during training often leads to an increase in output error and degradation of previously acquired knowledge, making the expansion process time-consuming and dependent on complex manual adjustments. The method developed in this research performs the parallel addition of a new hidden layer and a new output layer alongside the existing output layer. Unlike conventional approaches, the added layers create a new data flow toward the network output, forming an auxiliary processing branch in which the new layers learn without degrading previously acquired knowledge. As a result, the proposed insertion strategy collaborates with the existing layers, characterizing the method as collaborative. After insertion, an integration step is carried out in which the auxiliary branch and the original output layer are combined to consolidate the learning of the new layers. This ensures that the global network error exhibits a decreasing behavior while maintaining learning stability. At the end of the process, the previous output layer becomes part of the new hidden layer, forming a network with a single output layer. The proposed method was evaluated on different classification datasets, demonstrating its ability to preserve previously consolidated knowledge, consistently reduce output error, and maintain stability during network expansion. The results indicate that the method provides a robust alternative for the automatic expansion of adaptive neural networks, improving training time efficiency, reducing the need for manual adjustments, and promoting greater flexibility and reliability in classification and pattern recognition applications.