Question

The Projects of 'Introduction to Al' for Undergraduate Stu dents (2018-19) Both of the following two projects should be completed and submitted. Project 1. Finding a set of data with provided labels, such as the dataset for h andwritten digit recognition, face recognition, speech recognition, function re gression, and etc. Dividing the original dataset into training subset and test su bset and completing the following two experiments: 1) using the method of "neural networks with supervised learning" to pref orm the training on the training subset and validate the training results on the test subset. Analyzing the reasons behind the results of your experiments an d proposing the possible solutions to improve the performance of your metho d; 2) using the method of "unsupervised learning" to performing the trainin g on the training subset, i.e., ignoring the provided labels in the training. Analy zing the relationship between the unsupervised learning and the above super vised learning based on the results of them. You need to submit an experiment report to show the details of your exp eriments

Answer 1

# MNIST dataset is employed here as requested and code consists of in-built keywords defined by tensorflow and python which cannot be changed so don't modify the keywords

import tensorflow as tf

from tensorflow.examples.tutorials.mnist

import input_data

mnist = input_data.read_data_sets("/tmp/data",one_hot=True)

n_nodes_hl1 = 1000

n_nodes_hl2 = 1000

n_nodes_hl3 = 1000

n_classes = 10

batch_size = 100

x = tf.placeholder('float',[None,784])

y = tf.placeholder('float')

# Now let’s create the architecture

def neural_network_model(data):

  

    hidden_1_layer = {'weights' : tf.Variable(tf.random_normal([784,n_nodes_hl1])),

                     'biases' : tf.Variable(tf.random_normal([n_nodes_hl1]))}

    #bias is used to make some neurons fire even if all inputs is 0

    hidden_2_layer = {'weights' : tf.Variable(tf.random_normal([n_nodes_hl1,n_nodes_hl2])),

                     'biases' : tf.Variable(tf.random_normal([n_nodes_hl2]))}

    hidden_3_layer = {'weights' : tf.Variable(tf.random_normal([n_nodes_hl2,n_nodes_hl3])),

                     'biases' : tf.Variable(tf.random_normal([n_nodes_hl3]))}

    output_layer = {'weights' : tf.Variable(tf.random_normal([n_nodes_hl3,n_classes])),

                     'biases' : tf.Variable(tf.random_normal([n_classes]))}

    l1 = tf.add(tf.matmul(data,hidden_1_layer['weights']),hidden_1_layer['biases'])

    l1 = tf.nn.relu(l1)

    

    l2 = tf.add(tf.matmul(l1,hidden_2_layer['weights']),hidden_2_layer['biases'])

    l2 = tf.nn.relu(l2)

    

    l3 = tf.add(tf.matmul(l2,hidden_3_layer['weights']),hidden_3_layer['biases'])

    l3 = tf.nn.relu(l3)

    

    output = tf.matmul(l3,output_layer['weights'])+ output_layer['biases']

    

return output

# lets define the cost function and train it

def train_neural_network(x):

    prediction = neural_network_model(x)

    

    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction,labels=y))

    

   

    optimizer = tf.train.AdamOptimizer().minimize(cost)

   

    hm_epochs = 10

    

    with tf.Session() as sess:

        sess.run(tf.global_variables_initializer())

       

        for epoch in range(hm_epochs):

            epoch_loss = 0

            for _ in range(int(mnist.train.num_examples/batch_size)):

                epoch_x,epoch_y = mnist.train.next_batch(batch_size)

                _, c = sess.run([optimizer,cost], feed_dict = {x:epoch_x,y:epoch_y})

                epoch_loss += c

            print('Epoch',epoch,'Completed out of',hm_epochs,'loss:',epoch_loss)

                    

        

        correct = tf.equal(tf.argmax(prediction,1),tf.argmax(y,1))

        accuracy = tf.reduce_mean(tf.cast(correct,'float'))

        print('Accuracy:',accuracy.eval({x:mnist.test.images,y:mnist.test.labels}))

# Let’s execute it

train_neural_network(x)