Assignment 5 – Convolutional Neural Networks and MNIST with Tensorflow. solved

Description

Assignment overview. This assignment is designed to introduce you to the convolutional neural networks and their implementation with Tensorflow. The first question is just a basic implementation of a 2-dimensional convolution with the example of an edge detector. Your next task is to go through the basic Tensorflow tutorials provided by Google and then to experiment with some variations on the MNIST dataset.

Questions:

1. [10 marks] Implement an edge filter to filter an image. The convolution should be implemented with basic array processing without the use of a convolution function or an exciting edge filter. You need to look up how to import and image into a numpy array. Show the original and the filtered image within you Jupyder file.
2. [40 marks] Go through the Tensorflow tutorials, specifically https://www.tensorflow.org/get_started/premade_estimators and https://www.tensorflow.org/tutorials/layers. A slightly modified (simplified) version of the basic convolutional network for the MNIST data set is provided in the file test_MNIST_cnn.py. Get familiar with the code provided. Compare it to the code provided in the Tensorflow website https://github.com/tensorflow/tensorflow/blob/r1.6/tensorflow/examples/tutorials/layers/cnn_mnist.py

In the provided program you should be able to change the number of classes to speed up running in case your system is slow. Also, you may need to use smaller batches.

• [10 marks, 5 marks for Grads] Modify the code to have only one convolutional layer by removing the second convolutional layer. Save your program in the MNIST_conv1.ipynb.
• [10 marks, 5 marks for Grads] Modify the code MNIST_conv2.ipynb to have 3 convolutional layers. Add another convolutional layer with filter size (5 by 5) after the second convolutional layer. Save this code in MNIST_conv3.ipynb.
• [20 marks, 10 marks for Grads] Evaluate the performance of networks with 1, 2 or 3 layers and report on your findings.
• Grad Students only [20 marks] Add label-noise to the training data set and train the two-layer model. To make mislabeled data you can simply shuffle some of the labels. For example, if you want to make 10 percent of the data mislabeled, you can permute/shuffle %10 of the labels. Evaluate the performance of the network with different rates of label noise (%10, %25, %50, %75, %100) applied. Illustrate your result in a plot. Explain the performance of your network with %100 noise, specifically the difference between the training and test error.