Title:Computational Eco-Systems for Handwritten Digits Recognition

Abstract:On tasks such as the recognition of handwritten digits, traditional methods from machine learning and computer vision have always failed to beat human performance. Inspired by the importance of diversity in biological system, we built an heterogeneous system that could achieve this goal. Our architecture could be summarized in two basic steps. First, we generate a diverse set of classification hypothesis using both Convolutional Neural Networks, currently the state-of-the-art technique for this task, among with other traditional and innovative machine learning techniques. Then, we optimally combine them through Meta-Nets, a family of recently developed and performing ensemble methods. In the resulting parliament of classifiers all hypothesis, despite of their accuracy or methodology, are considered. On the very competitive MNIST handwriting benchmark, our method is the first to beat human performance with 0.17% error rate, surprisingly showing diversity to be the key for success in decision making.