Modeling and Training of Neural Processing Systems

Abstract

The field of deep learning has become more and more pervasive in the last years as we have seen varieties of problems being solved using neural processing techniques. Image analysis and detection, control, speech recognition, translation are only a few prominent examples tackled successfully by neural networks. Thereby, the discipline imposes a completely new problem solving paradigm requiring a rethinking of classical software development methods. The high demand for deep learning technology has led to a large amount of competing frameworks mostly having a Python interface - a quasi standard in the community. Although, existing tools often provide great flexibility and high performance, they still lack to deliver a completely domain oriented problem view. Furthermore, using neural networks as reusable building blocks with clear interfaces in productive systems is still a challenge. In this work we propose a domain specific modeling methodology tackling design, training, and integration of deep neural networks. Thereby, we distinguish between three main modeling concerns: architecture, training, and data. We integrate our methodology in a component-based modeling toolchain allowing one to employ and reuse neural networks in large software architectures.