Homework 03: Multilayer Perceptron for Multiclass Discrimination Solution

Description

In this homework, you will implement a multilayer perceptron for multiclass discrimination in R, Matlab, or Python. Here are the steps you need to follow:

1. Read Section 11.7.3 from the textbook.

2. You are given a multivariate classification data set, which contains 1000 clothing images of size 28 pixels × 28 pixels (i.e., 784 pixels). These images are from five distinct classes, namely, T-shirt, trouser, dress, sneaker, and bag. The figure below shows five sample figures from each class. You are given two data files:

1. 1. hw03_images.csv: letter images,

1. 2. hw03_labels.csv: corresponding image labels (1: T-shirt, 2: trouser, 3: dress,

4: sneaker, and 5: bag).

3. Divide the data set into two parts by assigning the first 500 images to the training set and the remaining 500 images to the test set.

3. Train a multilayer perceptron for multiclass discrimination using the sigmoid activation function for twenty nodes in the hidden layer ( = 20) and using the softmax activation function for five nodes in the output layer. You should develop a backpropagation algorithm under batch learning scenario with the following learning parameters.

eta <- 0.0005

epsilon <- 1e-3

max_iteration <- 500

You should initialize the weight parameters to the values given in initial_W.csv and initial_V.csv files before running your algorithm.

5. Draw the objective function values throughout the iterations. Your figure should be similar to the following figure.

What to submit: You need to submit your source code in a single file (.R file if you are using R,

.m file if you are using Matlab, or .py file if you are using Python) and a short report explaining your approach (.doc, .docx, or .pdf file). You will put these two files in a single zip file named as STUDENTID.zip, where STUDENTID should be replaced with your 7-digit student number.

How to submit: Submit the zip file you created to Blackboard. Please follow the exact style mentioned and do not send a zip file named as STUDENTID.zip. Submissions that do not follow these guidelines will not be graded.

Late submission policy: Late submissions will not be graded.

Cheating policy: Very similar submissions will not be graded.