FNN Web Demo: A Visualization Demo Platform for Feedforward Neural Networks Implemented from Scratch

FNN Web Demo is an educational project created by developer Khalil Burns, consisting of two core parts: a feedforward neural network (FNN) algorithm implemented from scratch, and a web demo platform built with Spring Boot. The project aims to address the difficulty in understanding abstract principles during neural network learning. By combining theory and practice, it provides learners with an interactive visualization and experimental environment to help them deeply understand the underlying working mechanisms of neural networks.

As the cornerstone of deep learning, the internal mechanisms of neural networks are abstract and difficult for beginners to understand. While modern frameworks (such as TensorFlow and PyTorch) simplify construction through high-level encapsulation, they mask the mathematical principles behind the algorithms. Implementing a neural network from scratch is the best way to understand its essence. This project combines theoretical learning with practical demonstrations, allowing learners to both read the underlying implementation code and observe the training process and prediction behavior through an interactive interface, which has significant educational value.

The core of the project is the FNN implementation that does not rely on existing deep learning libraries, covering:

1. Network Architecture: Manually define the structure of input, hidden, and output layers, and initialize weight matrices;
2. Activation Functions: May implement Sigmoid (for binary classification output), ReLU (to alleviate gradient vanishing), Tanh, etc.;
3. Forward Propagation: Handle matrix multiplication, bias addition, and application of activation functions to ensure correct data flow;
4. Backward Propagation: Apply the chain rule to calculate gradients from the output layer to the input layer, guiding parameter updates;
5. Loss Functions and Optimization: May implement Mean Squared Error (for regression), Cross-Entropy (for classification) losses, as well as SGD or its variant optimization algorithms.

The web platform is built with Spring Boot and adopts the MVC pattern:

• Controller Layer: Receives HTTP requests, calls business logic, and returns responses;
• Service Layer: Encapsulates core neural network logic (training, prediction, parameter management);
• RESTful API: Supports asynchronous communication with the frontend, such as creating networks, uploading data, starting training, etc.;
• Real-time Communication: May use WebSocket or long polling to push training progress, providing real-time feedback such as loss curves and accuracy.

The platform makes abstract processes intuitive through visualization:

1. Network Structure Visualization: Displays neuron connections and weight distribution (the depth of color indicates the size of weights);
2. Dynamic Training Charts: Real-time display of the changing trend of loss functions to help understand gradient descent convergence;
3. Decision Boundary Visualization: For 2D classification problems, shows the category regions divided by the network and the optimization process;
4. Interactive Experiments: Allows adjustment of hyperparameters (learning rate, hidden layer size, etc.) and immediate observation of effects to establish an intuitive understanding of hyperparameter impacts.

Learning Path:

1. Read and understand the FNN code implemented from scratch to build a solid foundation in mathematical principles;
2. Run the web platform and explore network behavior through interactive experiments (such as different datasets, hyperparameter tuning);
3. Extend the project (add activation functions, regularization, CNN extensions, etc.) to transform theory into practice.

Technical Gains: Master neural network mathematical principles, numerical calculation stability techniques, Java enterprise-level web development, front-end and back-end separation architecture, machine learning model deployment practices, etc.

As an open-source educational project, FNN Web Demo fills the gap between theory and practice in deep learning education (it is neither just abstract formulas nor just framework calls). Its open-source nature allows community collaboration for improvements (such as adding visualization types, complex architectures, sample datasets), enabling continuous evolution to benefit more learners. At the same time, it can serve as supporting experimental material for computer science courses to help students deepen their understanding, and it is also easy to explain neural network concepts to non-technical backgrounds, embodying the educational philosophy of 'learning by building'.