Date of Publication : 29th June 2024

Abstract:With the role of AI becoming more and more decisive in technological progress and human civilization's trajectory, we must look into this very critical limitation of Neural Networks – their inability to reason and comprehend mathematics. In this research, we investigate the mathematical capabilities of neural networks, focusing on symbolic integration—a pivotal mathematical task with diverse practical applications. We approach integration as neural translation task, using transformers. A novel digit encoding method has also been introduced, which significantly improves the performance of our basic integrator model. The study includes discussions on dataset generation, preprocessing procedures for handling symbolic mathematics data, detailed model architecture, and a comparative analysis of achieved results with mathematical software solutions.

Reference :