Create a .json file, e.g.:

./data/or_gate.json

  {
    "bias_enabled": false,
    "topology": [ 2, 3, 5, 1 ],
    "max_avg_sq_error": 0.001,
    "inputs": [
      [ 0, 0 ],
      [ 0, 1 ],
      [ 1, 0 ],
      [ 1, 1 ]
    ],
    "targets": [ 0, 1, 1, 1 ]
  }

Then run it as a neural net:

./bin/neural_net.rb ./data/or_gate.json

Download MNIST csv(s):

  https://pjreddie.com/media/files/mnist_train.csv
  https://pjreddie.com/media/files/mnist_test.csv

Convert MNIST data to a JSON formatted neural net spec:

↪ time ./bin/mnist_csv_to_json.rb pretrain 0.01 64 # quick -- see some quick progress
↪ time ./bin/mnist_csv_to_json.rb train # final - takes longer

PreTrain on MNIST data: (just to see some progress, on recognizing some digits 0 - 9): ./bin/neural_net.rb ./data/mnist_pretrain.json

Train on MNIST data:

↪ ./bin/neural_net.rb ./data/mnist_train.json

FIXME: this depends on being able to save the trained net?! Convert MNIST data to a JSON formatted neural net spec:

↪ time ./bin/mnist_csv_to_json.rb test

Test MNIST

↪ ./bin/neural_net.rb ./data/mnist_test.json

TODO:

update input to be closer to output, i.e.: ./output/initial_unbiased_or_gate.json

a) include (optional) connections & weights that way a trained net can be saved and restored for use later...

b) activation function specified (by name)

c) visualize output (not just JSON, but html/css OR canvas)