Transparent Deep Reasoning Neural Network

This project provides a deep neural network architecture with 20 fully connected layers, where each layer logs its internal activity during forward propagation. The logging includes the input tensor, output, layer weights (if applicable), and a feature-level influence vector computed as a linear product between input and weight matrix. The goal is to allow full layer-by-layer inspection of how information flows through the model, enabling a transparent reasoning trace without modifying the model’s original behavior. This approach supports reproducible introspection and explainability in deep networks by design, rather than through post hoc analysis. A sample reasoning trace is included, along with brief documentation and the core source code. This work is intended for researchers and developers working on interpretable AI, educational tools, and internal logic analysis of neural architectures. Автор: Бато Найданов Лицензия: CC BY-NC 4.0