MNIST Digit Recognition — V1
Java
Earlier educational feedforward neural network in pure Java (sigmoid activations + backpropagation) with MSE convergence checking and CSV weight serialization. Built without any external ML library, focused on the core mechanics of training and inference.
The first iteration of my from-scratch neural network work, built for the Artificial Intelligence course. No external libraries — just double[][] and the chain rule.
What’s inside
Neuron: weights, bias, sigmoid activation, and its derivative; handles forward propagation locally.Layer: a collection of neurons; manages forward propagation for the entire layer; initializes weights randomly.NeuralNetwork: orchestrates training and testing, implements backpropagation, computes MSE, and saves/loads trained weights to/from CSV.Main: entry point with three execution modes — full training, test-with-saved-weights, and an interactive console mode.
Training process
- Forward pass: compute the network output for each sample.
- Error: difference between predicted and expected.
- Backward pass: compute gradients via backpropagation through each layer.
- Update: adjust weights and biases via gradient descent.
- Convergence check: monitor MSE across epochs; stop when below threshold or when MSE starts climbing.
Defaults
| Parameter | Value |
|---|---|
| Learning rate | 0.1 |
| MSE threshold | 0.0001 |
| Architecture | 1 hidden layer |
| Training data | Lines 1–280 of dataset.csv |
| Test data | Lines 1–800 of dataset.csv |
Output artifacts
| File | Contents |
|---|---|
pesos.csv | Saved weights after training |
mse_values.txt | MSE per epoch — useful for plotting the learning curve |
How it relates to the V2 work
This V1 network laid the groundwork for the much more substantial MNIST Digit Recognition NN V2, replacing sigmoid with ReLU, adding a softmax output layer with cross-entropy loss, switching to mini-batch SGD, and shipping a live web playground that runs the trained weights directly in the browser.