White Box Network: Obtaining a right composition ordering of functions

Eun saem Lee, Hyung Ju Hwang

Keywords: white box, black box, function composition, neural network, ordering functions, reverse engineering, programmable logic controller, plc, white box network, WBN

TL;DR: We presented a new model called the WBN, which obtains the exact order and correct inputs of function blocks to compose them for constructing target functions.

Abstract: Neural networks have significantly benefitted real-world tasks. The universality of a neural network enables the approximation of any type of continuous functions. However, a neural network is regarded as a non-interpretable black box model, and this is fatal to reverse engineering as the main goal of reverse engineering is to reveal the structure or design of a target function instead of approximating it. Therefore, we propose a new type of a function constructing network, called the white box network. This network arranges function blocks to construct a target function to reveal its design. The network uses discretized layers, thus rendering the model interpretable without disordering the function blocks. Additionally, we introduce an end-to-end PathNet structure through this discretization by considering the function blocks as neural networks

Code: https://github.com/ekRmxskTek/WBN/