hthuwal / machine-learning-iitd

Assignments for the Machine Learning course (COL774) at IITD

mcs-ml-assignments

Assignments for the Machine Learning course (COL774) at IITD

Assignment 1

1. Linear Regression using Gradient Descent

• Leaning Rate(η) = 0.001

• Leaning Rate(η)= 0.017

2. Locally weighted Linear Regression using Normal Equations

• Bandwidth(τ) = 0.30

3. Logistic Regression using Newton's Update Method

4. Gaussian Discriminant Analysis

Assignment 2

1. Naive Bayes

• Basic
• +Stemming
• +Stopwords Removal
• +Feature Engineering

2. SVM

• MNIST Hand written digit recognition
• Pegasos Algorithm: 92.49% Test accuracy
• LibSVM Linear Kernel: 92.78
• Gaussian Kernel: 97.23

Assignment 3

1. Decision Trees and Random Forests

a. Decision Trees

b. Post Pruning Decision Trees

• Inefficient Implementation (Takes too much time)

c. Dynamically calculate median of Numerical Data

Instead of calculating the median of the numerical data beforehand median at a particular node is calculated dynamically using only the data coming to a particular node.

d. Scikit Learn Decision Tree and Random Forests

• Dtree Test acc: 84.80%
• Rforest Test acc: 85.485%

2. Neural Network

a. Implement Neural Network Class

• Forward Pass: to update outputs
• Backward Pass: backpropogation

b. Visualizing Decision Boundaries

• Logistic Regression

• Single Hidden Layer with 1 neuron

• Single Hidden Layer with 2 neurons

• Single Hidden Layer with 3 neurons

• Single Hidden Layer with 5 neurons

• Single Hidden Layer with 10 neurons

• Single Hidden Layer with 20 neurons

• Single Hidden Layer with 40 neurons

• Two Hidden Layers with 5 neurons each

c. MNIST Handwritten Digits Recognition - 6 & 8

• SVM Linear: 98.4722% Test accuracy
• Single Perceptron: 98.91% Test accuracy
• 100 units in hidden Layer: 99.1389% Test accuracy

Assignment 4

1. K-Means

Test Acc: 36.685%

2. PCA + SVM

Linear SVM: 69.842% Test Acc

Rbf SVM: 81.795% Test Acc

3. Neural Network

Single Hidden Layer with 1000 neurons: 77.952%

4. CNN

Test Acc: 83.162%

5. Kaggle Competition Link to competition

• Preprocessing

  Normalize data to zero mean and unit vaiance

• Modified VGG13 Architecture

  

  Difference from original vgg13

  1. Dropout Layer after every Max Pool Layer
  2. No Max Pool Layer in Last Conv2D block

• Hyperparameters

  • number_of_epochs: 10
  • batch_size: 64
  • Adam optimizer