Modular Design Patterns for Generative Neuro-Symbolic Systems
conference paper
Developing systems that are able to generate novel outputs is one of the dominating trends in current Artificial Intelligence (AI) research. Both capabilities and availability of such generative systems, in particular of so-called Large Language Models (LLMs), have been exploding in recent years. While Neuro-Symbolic generative models offer advantages over purely statistical generative models, it is currently difficult to compare the different ways in which the training, fine-tuning and usage of the growing variety of such approaches is carried out. In this work, we use the modular design patterns and Boxology language of van Bekkum et al for this purpose and extend those to enable the representation of generative models, specifically LLMs. These patterns provide a general language to describe, compare and understand the different architectures and methods used. Our main aim is to support better understanding of generative models as well as to support engineering of LLM-based systems. In order to demonstrate the usefulness of this approach, we explore generative Neuro-Symbolic architectures and approaches as use cases for these generative design patterns.
TNO Identifier
1002470
Source title
GeNeSy’24: First International Workshop on Generative Neuro-Symbolic AI, May 2024, Hersonissos, Crete, Greece