Implementando uma máquina virtual diferenciável mínima em redes neurais recorrentes

Abstract

Deep Learning techniques have achieved impressive results in many domains over the last few years. However, it’s still difficult to produce understandable models that clearly show the embedded logic behind the decision process while still having competitive performance. One step in this direction is the recent development of neural programmers. In this work, it’s proposed a very simple neural programmer with an extensible differentiable virtual machine that can be easily integrated in existing deep learning architectures, providing modules with more transparent reasoning to current models. At the same time it enables neural networks to learn to write and execute algorithm within the same training environment. Tests conducted with the proposed network suggests that it has the potential to induce algorithms even without any kind of special optimization and being competitive with current recurrent neural networks architectures.