NeuroGo: Neural Network Implementation in Go

A minimal neural network framework.

NeuroGo lets you build, train, and interactively explore feedforward neural networks with features like ReLU, Softmax, Dropout, Backpropagation, Class Weights, Early Stopping, and more.

Features

• Simple Architecture: Single hidden layer neural network with configurable layer sizes
• Efficient Training: Implements mini-batch gradient descent with momentum
• Learning Rate Adaptation: Dynamic learning rate adjustment based on training progress
• High Performance: Achieves 93.33% accuracy on the Iris dataset
• Interactive CLI: User-friendly command-line interface for model training and evaluation

Usage

The CLI provides an interactive menu with the following options:

1. Load dataset
2. Define architecture
3. Train model
4. Evaluate
5. Exit

Training Process

The model is trained on the Iris dataset with the following parameters:

• Input layer: 4 neurons (sepal length, sepal width, petal length, petal width)
• Hidden layer: 6 neurons with ReLU activation
• Output layer: 3 neurons with softmax activation (Setosa, Versicolor, Virginica)
• Learning rate: Starts at 0.005 with dynamic reduction
• Batch size: 4
• Epochs: 400

Performance Results

The model achieves the following performance on the Iris dataset:

• Test Accuracy: 93.33%
• Test Loss: 0.5214
• Training Time: ~400 epochs
• Confusion Matrix:

  Predicted      Setosa  Versicolor   Virginica
  Actual
  Setosa              5          0          0
  Versicolor          0         14          2
  Virginica           0          0          9
  

Sample Predictions

The model shows high confidence in predictions:

• Setosa samples: ~94-95% confidence
• Versicolor samples: ~46-59% confidence
• Virginica samples: ~55% confidence

Architecture

The current implementation uses a simple yet effective architecture:

• Input layer: 4 neurons (sepal length, sepal width, petal length, petal width)
• Hidden layer: 6 neurons with ReLU activation
• Output layer: 3 neurons with softmax activation (Setosa, Versicolor, Virginica)

Training Parameters

• Learning rate: 0.005 (initial) with dynamic reduction
• Batch size: 4
• Epochs: 400
• Patience: 8 epochs
• Learning rate reduction factor: 0.3

Current Implementation Status

✅ Completed Features

1. Core Neural Network

   • Feedforward network with ReLU + Softmax
   • Training loop with batching
   • Early stopping
   • Learning rate scheduling
   • L2 regularization
   • He initialization for weights

2. Training & Evaluation

   • Train/Test data separation
   • Predict support
   • Evaluation metrics
   • Confusion matrix generation

3. CLI Interface

   • Interactive command-line interface
   • Dataset loading and preprocessing
   • Model training and evaluation
   • Real-time training progress display

License

This project is licensed under the MIT License - see the LICENSE file for details.