/FL-Interactive-Game

FL-Interactive-Game: Interactive web game that teaches basic components of Federated Learning

Primary LanguageHTMLMIT LicenseMIT

🎮 Federated Learning Interactive Game

Non-IID FL Game Demo IID FL Game Demo

Checkout the game at: FL-Interactive-Game

An interactive web-based demonstration that makes learning about Federated Learning fun! Train models across multiple clients, experiment with different parameters, and watch your global model improve - all without sharing raw data.

🌟 Features

  • Interactive Client Management: Control 5 independent clients training on MNIST data
  • Real-time Visualization: Watch training progress with live accuracy plots
  • Configurable Parameters:
    • Per-client learning rates and epoch counts
    • Communication dropout probability
    • Data distribution settings (IID vs Non-IID)
  • Model Aggregation: Combine client models into a stronger global model
  • Responsive Design: Works seamlessly on both desktop and mobile devices

🎯 Learning Objectives

This game helps you understand:

  • How Federated Learning works in practice
  • Impact of different learning rates and training durations
  • Effects of communication dropouts in distributed learning
  • Differences between IID and Non-IID data distributions
  • Model aggregation and its effects on global performance

🎲 How to Play

  1. Choose your data distribution (IID or Non-IID)

    • IID: Balanced data
    • Non-IID: Chaos mode
    • Remember to refresh when switching - those sneaky old models like to hang around! 🔄

  2. Configure each client:

    • Adjust learning rates (0.00001 to 0.001)
    • Set training epochs (5 to 15)

  3. Set global parameters:

    • Adjust communication dropout probability
    • This simulates real-world network conditions

  4. Start training:

    • Train clients individually or all at once
    • Watch the accuracy plots evolve
    • Aggregate models to improve global performance

🔧 Technical Details

  • Built with TensorFlow.js
  • Uses MNIST dataset (preprocessed for FL setting)
  • Neural Network Architecture:
    • Input Layer: 784 neurons (28x28 images)
    • Hidden Layer: 64 neurons, ReLU activation
    • Output Layer: 10 neurons, Softmax activation

🏆 Challenge Goals

  1. Beat 90% accuracy on the global model
  2. Achieve consistent performance across all clients
  3. Maintain good performance with high dropout rates
  4. Master both IID and Non-IID scenarios

Some Known Issues

  • Model states persist between data distribution switches (refresh required)
  • Some mobile devices may experience performance lag with multiple clients
  • High learning rates can cause training instability