Problem 1: The MNIST data set is consisted of gray-scale images of hand-written digits. Each image has 28 28 i e . The e a e 10 c a e , ha i , digi 0, 1, 2, , 9. The e a e 60,000 training images and 10,000 test images. The goal is to recognize the digit on the image. Use multi-class logistic regression for the hand-written digit recognition task. Give the recognition accuracy rate, and show the confusion matrix. The following code segment is for your reference:
import tensorflow as tf import numpy as np mnist = tf.keras.datasets.mnist from sklearn.linear_model import LogisticRegression from sklearn.model_selection import train_test_split from sklearn.metrics import confusion_matrix
(x_traino, y_train),(x_testo, y_test) = mnist.load_data()
x_train = np.reshape(x_traino,(60000,28*28)) x_test = np.reshape(x_testo,(10000,28*28)) x_train, x_test = x_train / 255.0, x_test / 255.0
logreg = LogisticRegression(solver='saga', multi_class='multinomial',max_iter = 100,verbose=2)
Problem 2: Build a two-layer neural network for the hand-written digits recognition task with the MNIST data set. The hidden layer has 512 nodes, and adopts the ReLU activation function; the output layer has 10 nodes, and adopts the softmax activation function. Use the cross-entropy error function, and run 5 epochs. Give the recognition accuracy rate and show the confusion matrix, both for the test set.
Problem 3
Consider a neural network that has 𝐾 output nodes. The error function adopted at the output layer is the sumsquared-error cost function 𝐸 ∑𝐾= 𝑦 −𝑡 , where 𝑦 𝑦 𝐱 is the k-th output of the n-th data sample, and 𝑡 is the k-th target value of the n-th data sample. The output nodes adopt the sigmoid activation function.
a. Derive the math expression of 𝛿 for the k-th output node.
b. A local structure of this neural network is shown in Figure 1. Assume that the right-most layer in the figure is the output layer, and you have obtained 𝛿 , =1, 2, , 𝐾 in question a.
Figure 1.
Derive the math expression of 𝛿 for the j-th node in the middle layer in the figure, that is, one preceding layer of the output layer. Assume that the middle layer adopts the tanh function as the activation function ℎ. Use the result you obtained in question a.
