MNIST Neural Network

A neural network implemented from scratch in Go to classify MNIST digits.

Architecture

• Input Layer: 784 units.
• Hidden Dense Layer: 128 units (ReLU activation).
• Output Dense Layer: 10 units (Softmax activation).
• Loss Function: Cross-Entropy.

Getting Started

Installation

Clone the repository and ensure Go is installed on your system. It should be compatible with Go 1.23.4. Then run:

go run .

Alternatively, you can build an executable:

go build -o main . && ./main

Dataset Requirements

Place the MNIST dataset files in a directory named datasets. Ensure the following files are included:

• t10k-images-idx3-ubyte
• t10k-labels-idx1-ubyte
• train-images-idx3-ubyte
• train-labels-idx1-ubyte

Results

After training, the program shows:

• Training accuracy and loss per epoch.
• Validation accuracy and loss per epoch.