Designing a robust and accurate classifier is one of the most important goals in the machine learning society. This issue becomes crucial in the case of multi-class problems. In this research, a new architecture of cascaded classifiers is proposed to handle multi-class tasks. The stages of the proposed cascade are broken into some sub-stages; each contains a number of classifiers. Here, LogitBoost is used as the base classifier due to its low sensitivity to the noisy samples. To assess the proposed method, other cascade structures are implemented and eleven datasets derived from UCI repository are selected as the benchmark. Experimental results imply on the effectiveness of the proposed cascade approach compared to LogitBoost as one of the most successful parallel ensemble structure.