Towards Universal Modeling Language for Neural Networks

Effective modeling is essential in both system and software development, serving as a key method for facilitating understanding, guiding design, and enabling communication among stakeholders.However, traditional universal system modeling languages like UML and SysML fall short when it comes to neural network modeling, where the structure, training, and deployment processes demand more detailed and specialized representations.Conversely, domain-specific languages like Keras, TensorFlow, PyTorch, and tools like Netron and Deep Learning Studio are too closely tied to specific implementation environments.This creates a significant challenge: the need to develop a universal modeling language specifically for neural networks that is both sufficiently simple (requiring a description of around ten pages) and capable of providing a detailed description of neural networks and their management.The main contribution of this paper is the introduction of such a language, called UM1NN, along with a detailed description and its application demonstrated through two important use cases: describing GPT-2 and defining the fine-tuning of GPT-2 for Question-Answering.