This is a C# implementation of a simple two layer neural net on the MNIST data set done completely from scratch with no external libraries.
Default values of 3000 iterations and a batch size of 100 consistently gets around low 90% accuracy on the test set with minimal training time.
dotnet run <batch size> <iterations>
Accuracy: 0.2, Iteration: 0
Accuracy: 0.21, Iteration: 10
Accuracy: 0.26, Iteration: 20
Accuracy: 0.25, Iteration: 30
Accuracy: 0.39, Iteration: 40
...
Accuracy: 0.9, Iteration: 2970
Accuracy: 0.94, Iteration: 2980
Accuracy: 0.87, Iteration: 2990
Calculating accuracy on test data...
Test Data Accuracy: 0.9044
If no values are specified for batch size and iterations the program will default to 100 and 3000.