Executive Summary : | OrthDNNs are theoretically motivated by generalization analysis of modern DNNs, with the aim to find solution to properties of network weights, that guarantee better generalization. This work is an effort to develop the theoretical framework and show OrthDNN’s applicability in mainly classification tasks such as early fault diagnosis in electric drives, which is an important issue, especially in sectors such as EV, aerospace and marine. So, this research on orthogonal deep networks, based on deep neural networks, is a fast developing research topic, and a promising solution. Most importantly, this idea does not modify the existing structure / or add anything which would cause additional computational burden, which is of immense value in such applications. |