Graph-based Neuro-Symbolic Programming for Explainable LLM-based Agents
A Talk by Yoan Sallami and Laƫtitia Del Banio
About this Talk
Neuro-Symbolic programming combines the efficiency of neural networks with the precision of formal systems, offering explainability through an intermediate language.
In this masterclass, we'll shift the focus from static finite state machines to dynamic programs, highlighting the differences and advantages of graph program interpreter-based agents.
The masterclass is designed for data scientists and machine learning engineers with prior knowledge of graphs and programming.
Participants will learn how to create knowledge graphs, embed data structures, and program agent systems using Neuro-Symbolic Programming. By the end, they'll have a solid understanding of this technique and hands-on experience implementing it.
Description
Neuro-Symbolic programming has been the foundation of many successful AI systems, harmoniously merging the efficiency of neural networks with the precision of formal systems. One of the primary advantages of such systems is the explainability they offer. By utilizing an intermediate language comprehensible to humans, they provide transparency often lacking in the "black box" nature of neural networks.
Today's approach to Agent system is to represent them as static finite state machines, in this masterclass we will present a new paradigm around dynamic programs. We will emphasize the differences between static finite state machines based agents and agents as graph program interpreter. Specially around the dynamic calling of sub-programs to solve a problem as well as more experimental features around self-programming systems.
This masterclass is targeted towards data scientists and machine learning engineers who are already familiar with graphs and high-level programming. We will explore techniques around the creation of a knowledge graph and the embedding of such data structure. We'll then explore the programming of an Agent system.
By the end of this masterclass, participants will have a comprehensive understanding of how to leverage Neuro-Symbolic Programming to create sophisticated and adaptable agent systems. They will also gain hands-on experience in implementing these systems, equipping them with valuable skills for future AI projects.
Key Topics
- Knowledge Graph
- Neuro-Symbolic Programming
- LLM based Agent System
Target Audience
- Data Scientists and Machine Learning Engineers
- Data Engineering
- Machine Learning Engineers
- Data Analysts
- Managers of the above
Goals
Get an hands-on experience in programming a Neuro-Symbolic agent system based on a Turing complete DSL based on Cypher.
Session outline:
- Brief introduction on Neuro-Symbolic programming
- Brief introduction to LLM-based agentic systems.
- Brief introduction on long-term memory for agent systems
- Brief introduction on different prompting techniques.
Format
This class is very hands-on.
The beginning of the class will start in a lecture format, but will quickly move to hands-on coding exercises.
We will be working with Jupyter notebooks, Mistral API and HybridAGI python library.
Level
Beginner - Intermediate
Prerequisite Knowledge
Basic Python and programming knowledge