With limited success, artificial neural networks bring several disadvantages. These shortcomings are related to architectural selection (e.g., number of neurons, number of layers) which are dependent on the number of inputs and outputs and the complexity of the input-output relationship. Also, training methods may require additional neurons and layers, increasing the size of the network, and may lead to underfitting or overfitting, rending the network useless beyond the data used for training and testing. The design process becomes an academic exercise in numerical investigation resulting in an untrusted “black box” where the designer has no influence over what is being learned. In the end, because of the depth of complexity, it’s impossible to understand how conclusions were reached.
A system is needed with an architecture where the designer has control over what is being learned and thus provides inherent elucidation. This paper presents and discusses such a system architecture comprising a set of mathematical functions and logic gates lending transparency and explanation to applications based on artificial neural networks. A relatively simple example shows how the system architecture replaces the regression form of supervised learning to determine the aerodynamic rolling moment coefficient given aileron deflections, using 80% less data than required by traditional system identification methods. The paper concludes by discussing the implications this system architecture has on the other forms of machine and deep learning (classification and clustering), predictive and prescriptive analytics, and due to the inclusion of logic gates, quantum computation.