# Learning Algebraic Representation for Systematic Generalization in Abstract Reasoning

**Authors:** Chi Zhang, Sirui Xie, Baoxiong Jia, Ying Nian Wu, Song-Chun Zhu, Yixin Zhu  
**arXiv:** [2111.12990](https://arxiv.org/abs/2111.12990)  
**Date:** 2021-11-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.12990v1/

## Key points

- Debate between connectionist and classicist approaches in artificial intelligence
- Connectionist models show strong performance in specific tasks but may lack systematic generalization
- Interest in exploring alternative methods to enhance reasoning abilities
- Researchers including Chi Zhang, Sirui Xie, Baoxiong Jia, Ying Nian Wu, Song-Chun Zhu, and Yixin Zhu focus on abstract reasoning for improving systematic generalization
- Development of the ALgebra-Aware Neuro-Semi-Symbolic (ALANS) learner combining elements of both connectionist and classicist methodologies
- ALANS learner uses abstract algebra and representation theory for cognitive architectures
- Ability to induce hidden operators during reasoning process for predicting answers
- Incorporating algebraic treatment enhances systematic generalization compared to traditional connectionist models
- Algebraic representations learned by ALANS can be decoded through isomorphism for accurate answers
- Potential of hybrid approaches in enhancing cognitive capabilities and leveraging mathematical principles in AI research.

## Summary

In the realm of artificial intelligence, the debate between connectionist and classicist approaches to intelligence has long been a topic of discussion. While connectionist models have shown remarkable performance in specific tasks, there is a growing body of evidence suggesting that their superiority may falter when it comes to systematic generalization. This observation has sparked interest in exploring alternative methods to enhance reasoning abilities. In response to this challenge, a team of researchers including Chi Zhang, Sirui Xie, Baoxiong Jia, Ying Nian Wu, Song-Chun Zhu, and Yixin Zhu have delved into the realm of abstract reasoning with a focus on improving systematic generalization. Their work revolves around the development of a hybrid approach that combines elements of both connectionist and classicist methodologies. The researchers introduce the ALgebra-Aware Neuro-Semi-Symbolic (ALANS) learner, which is designed to tackle abstract spatial-temporal reasoning tasks such as Raven's Progressive Matrices (RPM). The ALANS learner draws inspiration from abstract algebra and representation theory to create a novel framework for cognitive architectures. It comprises a neural visual perception frontend that processes visual information through object-based representation and an algebraic abstract reasoning backend that transforms this information into an algebraic structure. One key innovation of the ALANS learner is its ability to induce hidden operators on-the-fly during the reasoning process. By predicting the answer's representation based on these induced operators, the system can select the most suitable solution from a set of choices. Through extensive experimentation, the researchers demonstrate that incorporating an algebraic treatment significantly enhances systematic generalization compared to traditional connectionist models. Moreover, they show that the algebraic representations learned by ALANS can be decoded through isomorphism to generate accurate answers. This breakthrough not only showcases the potential of hybrid approaches in enhancing cognitive capabilities but also highlights the importance of leveraging mathematical principles in AI research. Overall, this study sheds light on new avenues for advancing artificial intelligence through a fusion of connectionist and classicist paradigms.