Randomize all 60,000 samples each epoch to prevent last-class bias.
03 Machine learning · Feb-May 2025
MNIST Neural Network.
A handwritten digit classifier built from first principles in Java: initialization, forward propagation, ReLU, softmax, cross-entropy, backpropagation, gradient clipping, and serialization.
INPUT / 28×28 GRAYSCALE
Draw a digit
Draw here
Your strokes are downsampled into the same 784 normalized pixel values used by the Java Guess class.
OUTPUT / SOFTMAX
Loading weights
Prediction
-
CONFIDENCEWaiting for model
01784 pixelsNormalize 0-1
02128 neuronsWeighted sum + ReLU
0364 neuronsLearned features + ReLU
0410 logitsStable softmax
LIVE INFERENCE This demo loads weights exported from trained_brain.ser and runs the network forward pass in your browser.
01No libraries.
No libraries.
No black boxes.
The goal was not simply to classify digits. It was to understand every operation hidden behind a modern machine-learning API.
Each 28×28 image becomes 784 normalized inputs. Two hidden layers learn useful features through ReLU activation. Ten raw output logits are converted into probabilities with a numerically stable softmax, and the highest probability becomes the prediction.
TRAINING / 10 EPOCHS
How the model learns
Compute hidden activations and ten raw output logits.
Measure error with cross-entropy against the one-hot label.
Propagate output error backward through every weight.
Clamp gradients to ±0.1, then apply learning rate 0.01.
EPOCH 0 / 10LOSS --
Exploding gradients
Large unstable updates prevented useful learning.
FIX Xavier initialization plus gradient clipping at ±0.1.
Always predicting 9
Ordered training made the last class dominate recent updates.
FIX Fisher-Yates-style sample shuffling before every epoch.
60,000MNIST training images
109K+learned weights and biases
0machine-learning libraries