Train details

This page shows training details of the neural network that is used for recognition of symbols on the main page.

Train parameters

Hidden layers sizes
200
Input image edge size px
20
Pen thickness
12
Normalize input data
Generate extra inputs by rotation
10
Test data set size
10 %
Learning rounds
1
Batch size
0
Max iterations per batch
500
Min deriv. comp. max magnitude
0.001
Initialization seed
1735010885
Optimization algorithm
ImprovedRpropMinus
Regularization
0.5
Rprop initial step
0.01
Rprop max step
10
Rprop step up mult
1.2
Rprop step down mult
0.5

Train data

Train set contain images that are used to directly train the neural network. Images in the Test set are used for evaluation only; the neural network was not trained for them, thus, they measure how well the network generalizes.

Test set

Following image shows the test set which is randomly selected as 10% of input data images.

Legend

White
Correctly predicted.
Orange
Not correctly predicted but the correct answer was 2nd or 3rd.
Red
Not correctly predicted.
Number in the corner
Confidence (output value of the neuron).