Java Neural Network on MNIST

This project implements a simple feed-forward neural network in Java for MNIST digit classification. The model is built from scratch using custom matrix operations, Xavier weight initialization, sigmoid activations, binary cross-entropy loss, and manual backpropagation.

Model Architecture

Layer	Size	Activation
Input	784	-
Hidden 1	128	Sigmoid
Hidden 2	64	Sigmoid
Hidden 3	32	Sigmoid
Output	10	Sigmoid

Training Curves

Epoch vs Cost/Loss

Epoch vs Accuracy

Java Model Evaluation

The Java neural network achieved 96.96% accuracy on the MNIST test set.

Per-Class Metrics

Digit	TP	FP	TN	FN	Precision	Recall	F1-score
0	970	33	8987	10	0.9671	0.9898	0.9783
1	1120	17	8848	15	0.9850	0.9868	0.9859
2	1001	28	8940	31	0.9728	0.9700	0.9714
3	993	72	8918	17	0.9324	0.9832	0.9571
4	955	20	8998	27	0.9795	0.9725	0.9760
5	838	24	9084	54	0.9722	0.9395	0.9555
6	921	13	9029	37	0.9861	0.9614	0.9736
7	994	31	8941	34	0.9698	0.9669	0.9683
8	941	37	8989	33	0.9622	0.9661	0.9641
9	963	29	8962	46	0.9708	0.9544	0.9625

Notes

• Rows in the confusion matrix represent actual labels.
• Columns in the confusion matrix represent predicted labels.
• TP, FP, TN, and FN are calculated one digit at a time using a one-vs-rest interpretation.