Neural Network for Pattern Recognition

Exercise 1: With Trainining_data_set_input[0][0], let's decide outputs from the hidden layer. Before you start exercises, you should click 'Initialize Weights' above.
- NO_IMAGES_IN_TRAINING_DATA_SET
- NO_NODES_INPUT_LAYER
- NO_NODES_HIDDEN_LAYER
- NO_NODES_OUTPUT_LAYER
- THRESHOLD
- ALPHA
- Training_data_set_input[0][], Training_data_set_input[1][] // For 'A' and 'B' respectively
- Weights_input_hidden_layers[][]
- Weights_hidden_output_layers[][]
- Outputs_input_layer[] // Outputs from input layer
- Outputs_hidden_layer[] // Outputs from hidden layer
- Outputs_output_layer[] // Outputs from output layer
- Expected_outputs[0][], Expected_outputs[1][] // For 'A' and 'B' respectively
- Errors[]
- Deltas_output_layer[] // Deltas for output layer
- Deltas_hidden_layer[] // Deltas for hidden layer
<div id='tr1-output'></div> <script> function compute_outputs_from_input_layer(image) { // Convert the 1-D array of image to a matrix of NO_NODES_INPUT_LAYER x 1. for (var j = 0; j < NO_NODES_INPUT_LAYER; j++) Outputs_input_layer[j] = image[j]; } function compute_outputs_from_hidden_layer() { for (var j = 0; j < NO_NODES_HIDDEN_LAYER; j++) { // sum of weighted input var Xj = 0; ??? Xj -= THRESHOLD; // through activation function Outputs_hidden_layer[???] = ??? } } function activate(X) { return 1 / (1 + Math.exp(-X)); // Sigmoid function } compute_outputs_from_input_layer(Training_data_set_input[0][0]); // 'A' and image number 0 compute_outputs_from_hidden_layer(); $('#tr1-output').html(Outputs_hidden_layer.toString()); </script>

Exercise 2: Let's decide outputs from the output layer. This exercise should be done after Exercise 1.
- NO_IMAGES_IN_TRAINING_DATA_SET
- NO_NODES_INPUT_LAYER
- NO_NODES_HIDDEN_LAYER
- NO_NODES_OUTPUT_LAYER
- THRESHOLD
- ALPHA
- Training_data_set_input[0][], Training_data_set_input[1][] // For 'A' and 'B' respectively
- Weights_input_hidden_layers[][]
- Weights_hidden_output_layers[][]
- Outputs_input_layer[] // Outputs from input layer
- Outputs_hidden_layer[] // Outputs from hidden layer
- Outputs_output_layer[] // Outputs from output layer
- Expected_outputs[0][], Expected_outputs[1][] // For 'A' and 'B' respectively
- Errors[]
- Deltas_output_layer[]
- Deltas_hidden_layer
<div id='tr2-output'></div> <script> function compute_outputs_from_output_layer() { for (var j = 0; j < ???; j++) { var Xj = 0; ??? Xj -= ???; ??? } } compute_outputs_from_output_layer(); $('#tr2-output').html(Outputs_output_layer.toString()); </script>

Exercise 3: Let's decide errors. This exercise should be done after Exercise 2.
<div id='tr3-output'></div> <script> function decide_errors(d_number) { ??? // Decide errors and save them into Errors[] } decide_errors(0); // 'A' $('#tr3-output').html(Errors.toString()); </script>

Exercise 4: Let's decide deltas for the output layer. This exercise should be done after Exercise 3.
<div id='tr4-output'></div> <script> function decide_deltas_for_output_layer() { ???; // Deltas into Deltas_output_layer[] } decide_deltas_for_output_layer(); $('#tr4-output').html(Deltas_output_layer.toString()); </script>

Exercise 5: Let's decide deltas for the hidden layer.
<div id='tr5-output'></div> <script> function decide_deltas_for_hidden_layer() { for (var j = 0; j < ???; j++) { var Dj = 0; ??? Deltas_hidden_layer[j] = ???; } } decide_deltas_for_hidden_layer(); $('#tr5-output').html(Deltas_hidden_layer.toString()); </script>

Exercise 6: Let's update weights between hidden layer and output layer.
<div id='tr6-output'></div> <script> function update_weights_hidden_output_layers() { for (var j = 0; j < ???; j++) for (var k = 0; k < ???; k++) Weights_hidden_output_layers[j][k] = ???; } update_weights_hidden_output_layers(); $('#tr6-output').html(Weights_hidden_output_layers[0].toString()); </script>

Exercise 7: Let's update weights between input layer and hidden layer.
<div id='tr7-output'></div> <script> function update_weights_input_hidden_layers() { ??? } update_weights_input_hidden_layers(); $('#tr7-output').html(Weights_input_hidden_layers[0].toString()); </script>

Exercise 8: Let's train the NN with other two images.
<div id='tr8-output'></div> <script> function train_network(a_number, img_number) { compute_outputs_from_input_layer(Training_data_set_input[a_number][img_number]); compute_outputs_from_hidden_layer(); ??? decide_errors(a_number); // alphabet number ???; ??? ???; ??? } var alphabet_number = 0; // 'A' var image_number = 1; train_network(alphabet_number, image_number); alphabet_number = 0; image_number = 2; train_network(alphabet_number, image_number); $('#tr8-output').html(Outputs_output_layer.toString()); </script>

Exercise 9: Let's assume that the NN is fully trained. Let's find if a given image is 'A' or 'B', by observing the output values.
<div id='tr9-output'></div> <script> function query_nn(image) { compute_outputs_from_input_layer(image); ??? } query_nn(Training_data_set_input[0][0]); // Let's try with an image in the training data set. $('#tr9-output').html(Outputs_output_layer.toString()); </script>

Exercise for Neural Network for Pattern Recognition