The Architect: Interactive Deep Learning Visualization in Excel

A preprint on arXiv introduces a novel system named The Architect, which transforms Microsoft Excel into an interactive tool for visualizing deep learning mathematics. Users can create a neural network using a concise table, and the system produces a workbook that illustrates the complete forward pass, along with optional backward pass and parameter updates. Live spreadsheet formulas display computed values, while users can modify inputs such as weights, labels, and hyperparameters, enabling Excel's recalculation engine to update related computations dynamically. This system fills a void in current deep learning tools, which typically obscure numerical details or only present architecture diagrams and training summaries. The Architect reveals the model's complete arithmetic, allowing users to inspect matrices, activations, losses, gradients, and updates through editable spreadsheet regions. This method offers a distinctive intermediary for comprehending neural networks, connecting high-level concepts with low-level execution. The preprint can be found on arXiv under the identifier 2608.13572.