MNIST Handwritten Digit Classifier

INFO 6205 Team Project — Spring 2026

A feedforward neural network (perceptron) built from scratch in Java to classify handwritten digits from the MNIST dataset.

---

Results

Metric	Value
Test Accuracy	97.33%
Training Samples	60,000
Test Samples	10,000
Epochs	5
Learning Rate	0.01

---

Confusion Matrix

Confusion Matrix (row=actual, col=predicted):
        0     1     2     3     4     5     6     7     8     9
  0   971     1     0     1     1     1     2     0     2     1
  1     0  1124     2     1     1     2     2     0     3     0
  2     3     5  1002     2     5     0     4     2     8     1
  3     1     0     6   979     0    16     0     2     2     4
  4     0     0     3     0   971     0     3     0     0     5
  5     3     0     0     2     2   879     2     1     2     1
  6     5     3     1     1     5    14   927     0     2     0
  7     2     7    17     5     7     1     1   956     2    30
  8     4     0     3     2     4     9     1     2   942     7
  9     0     3     0     2    12     6     1     2     1   982

---

Network Architecture

Input Layer        Hidden Layer       Output Layer
(784 neurons)  →  (128 neurons)   →  (10 neurons)
 28×28 pixels      ReLU activation    Softmax activation
                                      Classes: 0–9

• Input Layer: 784 neurons (one per pixel of 28×28 grayscale image), normalized to [0, 1]
• Hidden Layer: 128 neurons with ReLU activation, Xavier weight initialization
• Output Layer: 10 neurons with Softmax activation (one per digit class)
• Loss Function: Cross-Entropy Loss
• Optimizer: Stochastic Gradient Descent (SGD)

---

Project Structure

mnist-perceptron/
├── src/
│   ├── main/
│   │   ├── java/com/mnist/
│   │   │   ├── NeuralNetwork.java     # Core neural network (forward/backprop)
│   │   │   ├── MnistLoader.java       # MNIST data loader (.gz files)
│   │   │   └── Main.java              # Training + evaluation + confusion matrix
│   │   └── resources/
│   │       ├── train-images-idx3-ubyte.gz
│   │       ├── train-labels-idx1-ubyte.gz
│   │       ├── t10k-images-idx3-ubyte.gz
│   │       └── t10k-labels-idx1-ubyte.gz
│   └── test/
│       └── java/com/mnist/
│           └── NeuralNetworkTest.java # Unit tests (5 tests, all passing)
└── pom.xml

---

Unit Tests

All 5 unit tests pass using java.util.Random for input generation:

Test	Description	Status
testOutputSize	Output layer has exactly 10 neurons	✅
testSoftmaxSumsToOne	Softmax probabilities sum to 1.0	✅
testPredictInRange	Predictions are always in range [0, 9]	✅
testTrainingReducesLoss	Loss decreases after 200 training steps	✅
testWeightDimensions	Weight matrices have correct dimensions	✅

---

How to Run

Prerequisites

• Java 17+
• Maven 3.6+
• MNIST dataset files in src/main/resources/

Run Unit Tests

mvn test

Run Training & Evaluation

mvn compile exec:java -Dexec.mainClass="com.mnist.Main"

---

Data

MNIST dataset: 70,000 handwritten digit images (28×28 grayscale pixels)

• 60,000 training samples
• 10,000 test samples
• Source: http://yann.lecun.com/exdb/mnist/

---

Dependencies

• Java 17
• JUnit Jupiter 5.10.2 (unit testing)
• Maven Surefire Plugin 3.2.5