How i can improve my Neural Network for printed lettters recognition?

Hi all. I want to make a neural network for printed letters recognition. The type of NN is multilayer perceptron with 1 hidden layer. I used this dataset: https://archive.ics.uci.edu/ml/datasets/Letter+Recognition

I make my NN, but the testing result indicates that NN obtained 5% prediction accuracy. How I can make her more effective? I can't find other dataset for printed letters, to test with her. Of course I might have made a mistake somewhere and would be very grateful if you correct. This is code for my NN - 16 inputs and 26 targets ( I used 16000 simples for train and 4000 for test ).

clear all;
clc;
%reading from dataset
fid = fopen('letter-recognition.data', 'rt'); 
num_attrib = 16;
fmt = ['%s', repmat('%f', 1, num_attrib)];
datacell = textscan(fid, fmt, 'Delimiter', ',', 'CollectOutput', 1);
fclose(fid);
which_letter = char(datacell{1});
attribs = datacell{2};
target_codes = which_letter - 'A' + 1; %makes letters in numbers
%inputs
train_set = attribs(1:end-4000, :);
inputs = train_set.';% inputs 16x16000
%targets
train_targets = target_codes(1:end-4000);
letters = train_targets.';
targets = full(ind2vec(letters));% targets 16x16000 (eye)
%net
net = newff(minmax(inputs),[30 26], {'logsig','logsig'},'traingd');
net.inputweights{1,1}.initfcn = 'rands';
 net.LW{2,1} = net.LW{2,1}*0.1;
 net.b{2} = net.b{2}*0.1;
net.performFcn = 'mse';
net.trainParam.goal = 1e-2;
net.trainParam.show = 1;
net.trainparam.epochs = 800;
net.trainParam.lr = 0.1;
net.trainParam.max_fail = 5;
%divide data
 net.divideFcn = 'dividerand';  % Divide data randomly
net.divideMode = 'sample';  % Divide up every sample
net.divideParam.trainRatio = 70/100; % 70 % for train
net.divideParam.valRatio = 15/100;  % 15 % for validation
net.divideParam.testRatio = 15/100;  %15 % for  test
%train
[net,tr] = train(net,inputs,targets)
%prepare for sim
sim_attribs = attribs(end-3999:end, :);
sim_inputs = sim_attribs.';
sim_targets = target_codes(end-3999:end);
sim_letters = sim_targets.';
test_targets = full(ind2vec(sim_letters));
%sim
result = sim(net,sim_inputs);
%test NN
true = 0;
false = 0;
for index = 1:1:4000
    [M,I] = max(result( : ,index));
    [N,J] = max(test_targets( : ,index));
    if(I == J)
        true = true + 1;
    else
        false = false + 1;
    end
end
p_true = (true / 4000)*100 % how % true
p_false = (false / 4000)*100 % how % false