# Continual Reasoning: Non-Monotonic Reasoning in Neurosymbolic AI using Continual Learning

**Authors:** Sofoklis Kyriakopoulos, Artur S. d'Avila Garcez  
**arXiv:** [2305.02171](https://arxiv.org/abs/2305.02171)  
**Date:** 2023-05-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.02171v1/

## Key points

- Continual Reasoning methodology combines neural-symbolic systems with continual learning techniques
- Logic Tensor Networks (LTN) are used to simulate First-Order Logic (FOL) and handle deep learning and reasoning tasks
- Continual learning is incorporated through a curriculum of learning from knowledge and data with recall
- The combination addresses difficulties faced by neural-symbolic models in non-monotonic tasks
- Experimental results show significant improvement in accuracy compared to single-stage curriculum of learning
- Potential applications of Continual Reasoning on larger datasets and lifelong learning challenges are discussed
- Further exploration of advanced continual learning techniques and application to other neural-symbolic models is suggested
- Integration with software frameworks like PyReason could provide efficient implementation options

## Summary

This paper presents a novel methodology called Continual Reasoning that combines neural-symbolic systems with continual learning techniques to improve deep learning's performance in non-monotonic reasoning tasks. Specifically, Logic Tensor Networks (LTN) are chosen for their ability to simulate First-Order Logic (FOL) and effectively handle both deep learning and reasoning tasks. Continual learning is incorporated into LTNs through a curriculum of learning from knowledge and data with recall. This combination aims to address the difficulties faced by many neural-symbolic models when dealing with non-monotonic tasks. The authors apply Continual Reasoning to various reasoning examples including a prototypical non-monotonic reasoning problem, statistical relational reasoning tasks, and a Natural Language Understanding (NLU) task containing non-monotonic reasoning. Experimental results show significant improvement in accuracy compared to a single-stage curriculum of learning. The paper concludes by discussing the potential applications of Continual Reasoning on larger datasets and addressing lifelong learning challenges. The authors suggest further exploration of advanced continual learning techniques and the application of Continual Reasoning methodology to other neural-symbolic models. Integration with software frameworks like PyReason could provide efficient ways to implement this methodology. In summary, this paper demonstrates the effectiveness of Continual Reasoning compared to traditional approaches, highlighting its potential for future research and application in various domains.