This repository hosts the implementation of Neural Algorithmic Reasoning (NAR), an innovative approach blending classical algorithmic reasoning with the flexibility of neural networks.
The core focus lies in utilizing the NAR architecture to emulate the functionality of Breadth First Search (BFS) algorithm.
- Implementation of NAR architecture structured around the encode-process-decode paradigm.
- Integration of Graph Neural Networks (GNNs), specifically Message Passing Neural Networks (MPNN), to handle graph-structured data efficiently.
- Evaluation of NAR's efficacy using synthetic datasets generated through Erdos-Renyi and Barabasi-Albert models.
- Model: Contains the implementation of the NAR network architecture along with its constituent elements (Encoder-Processor-Decoder).
- generator: Houses a synthetic dataset generator designed to produce datasets for both training and testing the NAR model.