Bridging Logic and Learning: A Neural-Symbolic Approach for Enhanced Reasoning in Neural Models (ASPER)

Overview

This repository hosts the research work titled "Bridging Logic and Learning: A Neural-Symbolic Approach for Enhanced Reasoning in Neural Models (ASPER)." Under the acronym ASPEr (Answer Set Programming Enhanced Reasoning), it integrates Answer Set Programming (ASP) with a neural model for enhanced logical reasoning and problem-solving capabilities.

Contributions

Our research contributes to the field in several key areas:

• Innovative Integration: Development of a unique integration of ASP solvers with deep learning models, enhancing reasoning in tasks requiring logical consistency.
• Customized Loss Function: Introduction of a specialized loss function that combines standard loss with ASP solver outputs for improved learning efficiency.
• Practical Applicability: Demonstration of practical applicability and effectiveness through a detailed case study on solving Sudoku puzzles.
• Model Interpretability: Emphasis on model interpretability and adaptability across various problem domains.

Sudoku Puzzle Case Study

The methodology is applied to solve Sudoku puzzles, demonstrating significant improvements in model accuracy and reasoning capabilities.

Evaluation Results

The following figure illustrates our model's performance across different loss combinations in solving Sudoku puzzles:

Data Representation

The table below provides a detailed representation of the Sudoku puzzles, highlighting the model's accuracy:

Usage

• Download the ASPEN.ipynb notebook from the repository.
• Open the notebook in Google Colab for optimal performance.
• Execute the cells in ASPEN.ipynb in sequential order. Note: Running the cells sequentially is crucial for the correct execution of the notebook. Google Colab is recommended due to its pre-configured environment and easy access to necessary libraries."

Citation

If you find our research useful, please consider citing our paper:

@misc{machot2023bridging,
  title={Bridging Logic and Learning: A Neural-Symbolic Approach for Enhanced Reasoning in Neural Models (ASPER)},
  author={Fadi Al Machot},
  year={2023},
  eprint={2312.11651},
  archivePrefix={arXiv},
  primaryClass={cs.AI}
}