Description
In this programming assignment, you will implement techniques to learn a deep (two hidden layers)
neural network (NN) for image digit classification on the MNIST dataset
(https://en.wikipedia.org/wiki/MNIST_database). The NN will take an image of a hand-written
numerical digit as input and classify it into one of 10 classes corresponding to digits 0 to 9 respectively.
Given the training data, you will follow the equations in the lecture notes to train the NN using the back
propagation method and then evaluate its performance on the given testing data. The NN has one input
layer (784 nodes), two hidden layers with 100 nodes for each layer, and one output layer with 10 nodes,
corresponding to the 10 digit classes respectively. For training, you will learn the parameters Θ=[W1,
W1,0, W2, W2,0, W3,W3,0] by minimizing the squared loss for the output nodes, where W1 and W2 are the
weight matrices for the hidden layers, W3 is the weight matrix for the output layer, and W1,0 ,W2,0, W3,0 are
the corresponding bias vectors.
Specifically, given the training data D={x[m], y[m]}, m=1,2, …, M, where x[m] is a grayscale image of
28×28 (a 784×1 vector) and y[m] is output vector that follows the 1 of K encoding, with kth element of
being 1 and the rest being 0. Use ReLU activation function for the hidden nodes and the softmax (multiclass sigmoid) function for the output. You will implement the following methods to learn the
parameters Θ
1. Implement in Tensorflow the back propagation method to solve for Θ iteratively using all data in D.
Each time, randomly choose a mini-batch of 50 data from D. For each point in the mini-batch, perform
forward propagation and back propagation to compute the gradients of weight matrix and weight bias
vector for each layer. Then update the weights and bias using the average gradient of the gradients
computed from all the 50 samples. Initialize weight matrix to small different values and bias vectors all
to zero. Iteratively update each W with an appropriate learning rate until convergence. Save Θ using
pickle.dump functions (see below for details).
2. Using the given testing dataset T, evaluate the performance of your trained NN by computing the
classification error for each digit as well as the average classification errors for all digits. The
classification error for each digit is computed as the ratio of incorrect classification to the total number
images for that digit. Plot the average training classification error, average testing classification error, and
value of the loss function after each parameters update.
3. Do not use Tensorflow’s back propagation function. You however may use it to verify your
implementation. Submit your Tensorflow code via LMS, along with the required classification error plots,
and the saved parameters Θ .
Data format
The training data D contains 50000 training images (00001.jpg – 50000.jpg) in train_data folder, and
their labels are stored in train_label.txt with each raw being the label of the corresponding image.
The testing data T contains 10000 testing images (00001.jpg – 10000.jpg) in test_data folder, and their
labels are stored in test_label.txt.
Load images and normalize
Load the images in the order
Convert it to vector and normalize it to [0,1] by dividing the intensity of each pixel by 255
The following link teaches how to load images:
http://learningtensorflow.com/lesson3/
Output
Use the code below to save the learned parameters Theta in the following format
Θ= [W1, W1,0, W2, W2,0, W3, W3,0]
Θ is a list objects. W1 and W2 are the matrix of the weights for the hidden layers, W3 is the matrix of the
weights for the output layer, and W1,0, W2,0, W3,0 are the corresponding bias.
——————–
import pickle
filehandler = open(“nn_parameters.txt”,”wb”)
pickle.dump(Theta, filehandler, protocol=2)
filehandler.close()
—————–
Batch computing
To speed up the training, you can simultaneously compute the gradients of 50 samples at the
same time using 3-D tensor. For example, given two tensors AN*p*q and B N*q*m, tf.matmul(A,B) will
compute matrix multiplication N time and return a tensor with shape N*p*m. More details and
example can be found at:
https://www.tensorflow.org/api_docs/python/tf/matmul
