Proposal of a declarative and parallelizable artificial neural network using the notification-oriented paradigm

Since the 1960s, artificial neural networks (ANNs) have been implemented and applied in various areas of knowledge. Most of these implementations had their development guided by imperative programming (IP), usually resulting in highly coupled programs. Thus, even though intrinsically parallel in theory, ANNs do not easily take an effective distribution on multiple processors when developed under IP. As an alternative, the notification-oriented paradigm (NOP) emerges as a new programming technique. NOP facilitates the development of decoupled and distributed systems, using abstraction of knowledge through logical-causal rules, as well as the generation of an optimized code. Both features are possible by means of a notification-oriented inference process, which avoids structural and temporal redundancies in the logic-causal evaluations. These advantages are relevant to systems that have parts decoupled in order to run in parallel, such as ANN. In this sense, this work presents the development of a multilayer perceptron ANN using backpropagation training algorithm based on the concepts of a NOP implementation. Such implementation allows, transparently from high-level programming, parallel code generation that runs on multicore platforms. Furthermore, the solution based on NOP, when compared against the equivalent on IP, presents a high level of decoupling and explicit use of logic-causal elements, which are, respectively, useful to distribution, understanding and improvement of the application.