java-mnist-tutorial
Neural network to classify MNIST handwritten digits - implemented in Java
Data model
See MNIST for description of the MNIST data
Getting Started
mvn clean install -U
Prerequisites
Java 8
Description
To run the model execute Application.java
Code-base built as Spring Boot app
Configuration is set in AppConfiguration.java and application.yml
Main model logic (i.e. back-propagation) is in Processor.java
Training can be executed in 2 modes - general and concurrent. To change between modes edit application.yml entry:
spring:
profiles:
include: general
general - execute training in serial mode - each batch will be processed synchronously
concurrent - execute batches in concurrent mode - each batch will be processed concurrently in a thread-pool
Performance
Model produces ~94% test accuracy after 20 epochs
Epoch 1: Correct guesses 9191 of 10000
Epoch 2: Correct guesses 9251 of 10000
Epoch 3: Correct guesses 9301 of 10000
...
Epoch 19: Correct guesses 9400 of 10000
Epoch 20: Correct guesses 9427 of 10000