# NeuPSL: Neural Probabilistic Soft Logic

**Authors:** Connor Pryor, Charles Dickens, Eriq Augustine, Alon Albalak, William Wang, Lise Getoor  
**arXiv:** [2205.14268](https://arxiv.org/abs/2205.14268)  
**Date:** 2022-05-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.14268v3/

## Key points

- Authors introduce NeuPSL: Neural Probabilistic Soft Logic
- NeuPSL combines symbolic reasoning with deep neural networks
- Framework models interface between neural and symbolic representations
- Demonstrates seamless integration of neural and symbolic parameter learning and inference within NeuPSL
- Empirical evaluation shows over 30% improvement compared to independent neural network models
- NeuPSL outperforms existing NeSy approaches by up to 10% on MNIST-Addition task
- Achieves remarkable speedup of up to 40 times in canonical citation network task

## Summary

In their paper titled "NeuPSL: Neural Probabilistic Soft Logic," authors Connor Pryor, Charles Dickens, Eriq Augustine, Alon Albalak, William Wang, and Lise Getoor introduce a groundbreaking  that combines cutting-edge symbolic reasoning with the deep neural network's low-level perception. This innovative approach proposes  to model the interface between neural and symbolic representations, showcasing their versatility in encompassing NeuPSL and various other NeSy approaches. By leveraging this framework, the authors demonstrate how neural and symbolic parameter learning and inference can be seamlessly integrated within NeuPSL. Through an extensive empirical evaluation, the study highlights the advantages of employing NeSy methods, showcasing significant improvements of over 30% compared to independent neural network models. Specifically, on the MNIST-Addition task—a well-established NeSy benchmark—NeuPSL showcases its joint reasoning capabilities by surpassing existing NeSy approaches by up to 10% in scenarios with limited data availability. Additionally, in a canonical citation network task while achieving up to a remarkable 40 times speedup. Overall, The results presented in this paper not only demonstrate superior performance but also pave the way for enhanced integration of neural networks and symbolic reasoning in various applications within machine learning and artificial intelligence domains.