The study investigates the use of machine learning algorithms for multiclass classification using a framework from wireless communications. Previous works have presented and applied, to some extend, the idea of redundancy and error correction in classification tasks with learning algorithms. These are called error correction output codes (ECOCs). One of the core ideas of these approaches was to divide the multiclass classification problem into several binary problems trained separately, so that a well suited code design could be used instead of the classical one-hot mapping. By enhancing the Hamming distance between class labels, these works have increased the reliability of the classifier after applying a loss based decoder. Our work aims at extending these analogies and presenting new possibilities for the use of ECOCs. One of the main contributions so far is the creation of a blockwise loss function, applied to the output layers as LLRs. In addition, this new approach may be scalable for a very high number of classes. So far, results are promising with smaller datasets.