You can run the whole script with "./doit --download-mnist" or just "./doit.sh" if you already have the MNIST database in the current directory. You should get the following output (this takes a few hours on a 1.2Ghz Pentium-M): ---------------------------------------------------------------------- Loading the data file ... done. Database contains 60000 images of resolution 28x28 divided into 10 objects. Creating a new network (layers of sizes 784 200 10). Training the network with 20000 training and 20000 validation examples. 0 TRAINING 12235.8 (8.58%) TESTING 13030 (9.405%) 1 TRAINING 8839.31 (6.69%) TESTING 10132.9 (7.71%) 2 TRAINING 6502.38 (4.575%) TESTING 8268.75 (6.235%) 3 TRAINING 5656.71 (3.975%) TESTING 7637.6 (5.75%) 4 TRAINING 5456.68 (3.56%) TESTING 7683.5 (5.6%) [1] 5 TRAINING 4167.26 (2.64%) TESTING 6557.84 (4.82%) 6 TRAINING 4320.34 (2.7%) TESTING 6796.09 (4.89%) [2] 7 TRAINING 3725.38 (2.435%) TESTING 6307.99 (4.52%) 8 TRAINING 3946.58 (2.49%) TESTING 6614.6 (4.53%) [3] 9 TRAINING 3773.16 (2.24%) TESTING 6698.8 (4.67%) [4] 10 TRAINING 3485.74 (2.13%) TESTING 6539.64 (4.54%) 11 TRAINING 5903.21 (3.53%) TESTING 8881.58 (5.905%) [5] 12 TRAINING 3165.84 (1.89%) TESTING 6366.87 (4.385%) 13 TRAINING 3288.64 (2%) TESTING 6520.78 (4.5%) [6] 14 TRAINING 2849.94 (1.615%) TESTING 6201.43 (4.215%) 15 TRAINING 2693.19 (1.555%) TESTING 5991.35 (4.235%) 16 TRAINING 2827.86 (1.575%) TESTING 6181.83 (4.235%) [7] 17 TRAINING 2374.73 (1.355%) TESTING 5668.65 (3.77%) 18 TRAINING 2194.12 (1.255%) TESTING 5572.82 (3.705%) 19 TRAINING 2114.23 (1.155%) TESTING 5587.74 (3.71%) [8] 20 TRAINING 1909.78 (1.15%) TESTING 5377.6 (3.64%) 21 TRAINING 3064.62 (1.705%) TESTING 6642.78 (4.36%) [9] 22 TRAINING 1832.23 (1.04%) TESTING 5386.12 (3.575%) 23 TRAINING 1695.47 (0.95%) TESTING 5342.3 (3.61%) 24 TRAINING 1699.28 (0.935%) TESTING 5331.84 (3.46%) 25 TRAINING 1478.36 (0.835%) TESTING 5075.09 (3.335%) 26 TRAINING 1528.62 (0.865%) TESTING 5221.1 (3.41%) [10] Saving network simple.mlp ... done. Loading the data file ... done. Database contains 10000 images of resolution 28x28 divided into 10 objects. Loading network simple.mlp ... done (layers of sizes 784 200 10) Error rate 2599.54 (3.42%) ---------------------------------------------------------------------- The computation produces a file simple.mlp containing the learnt perceptron.