PGPAIML - Program Outline

The post-graduate program in Artificial Intelligence & Machine Learning (PGPAIML) is a

12-month, online, comprehensive program covering various facets of analytics –
including statistical models, machine learning and deep learning – all in an attempt to
help businesses make smart decisions, and build more intelligent products.

The program spans over 400 hours of learning sessions over the course of 12 months.
Quizzes, online instruction, labs, projects and hackathons are all completed online. The
program comprises of the following courses:

1 Introduction to Python
2 Applied Statistics
3 Supervised Learning
4 Ensemble techniques
5 Unsupervised Learning
Feature selection, Model Selection &
6
Tuning
7 Recommendation Systems
Introduction to Neural Networks &
8
Deep Learning
9 Computer Vision
10 Visualization using Tensor board
11 Natural Language Processing
14 Reinforcement Learning
15 Introduction to GANs
+ Industry Case Sessions
+ Capstone Project
+ Hackathons

Confidential – Not to be shared without

permission
Please Note - ​This is subjected to change as per the requirements.

Learning model
The program is divided into two parts: Conceptual Learning and Application.
Logistically, the program delivers on both these outcomes regularly. A typical week
consists of reviewing at least 3 to 4 hours of recorded videos concluded by a
dedicated 2-hours mentor session conducted by Industry experts that focuses on
clarifying doubts from the students. Quizzes are conducted at the end of each week to
reinforce the concepts and projects at the end of every course to emphasize the
learning through practice. Submission of projects is done online through Olympus, our
Learning Management System.

Attending all mentored sessions is mandatory for all participants. Students are required
to score a minimum of 60 out of 100 in every course and successfully completed the
Capstone project.

AIML – Course Curriculum

Introduction to Python
● Installation of Python
● Numpy
o Basics of Numpy array
o Broadcasting an array
o Matrix indexing
o Selection techniques
o Saving and loading arrays
● Pandas
o Series
o Indexing elements of a series
o Dictionaries

Confidential – Not to be shared without

permission
 o Data frame
o Different ways of indexing in a data frame
o Conditional indexing in a data frame
o Dropna and fillna
o Groupby, merging similar to SQL logic
o Filtering, sorting and indexing
o Loops and functions
o Saving and loading a csv, excel file

● Seaborn for data visualization

o Univariate analysis – Dist plot, count plot, Boxplot, Bar chart
o Bivariate analysis – Pair plot, Reg plot, Joint plot, Point plot, Factor plot,
Strip plot, Swarm plot
● Case Study

Statistical Learning
● Introduction to Statistical analysis
o Why statistics?
o Popular statistical methods in the field of AI & ML
o Types of statistics – Descriptive & Inferential
o Common terminologies in Statistics
o Data sources – Primary and Secondary
o Types of data – Qualitative and Quantitative
● Descriptive Statistics
o Histogram
o Frequency distribution
o Measures of central tendency (3M) – Mean, median, mode
o Measures of dispersion – Range, IQR, SD, variance & coefficient of
variation
o Empirical and Chebyshev rule
o Five number summary​, ​QQ plots, Quantile plot, scatterplot.
o Visualization – Scatter plot, correlation analysis
● Introduction to probability
o Basic probability: Definition and examples
▪ Mutually exclusive events

Confidential – Not to be shared without

permission
 ▪ Independent events
▪ Rules for computing probability
o Conditional probability
o Baye’s theorem
o Applications of Bayes theorem in real life scenario
o Probability distributions
▪ Binomial distribution
▪ Poisson distribution
▪ Normal distribution
● Hypothesis testing & scores
o Concept of sampling distribution
o Central Limit theorem
o Formulation of hypothesis – Null and Alternate
o Type I and II error
o Hypothesis testing roadmap
o Type of Hypothesis – One tail vs Two tailed test
▪ Confidence Interval
▪ Single sample - Z test
▪ Single sample – T test
o Two sample test
▪ Paired T test
o Z test of proportion
o Chi-square test of variance
o F-ratio test of variance
o ANOVA

SUPERVISED LEARNING
● Introduction to machine learning
o What is machine learning?
o When is machine learning useful?
o Application of machine learning in different domains
o Categories of machine learning
▪ Supervised learning – concepts, preparing data, training and
testing

Confidential – Not to be shared without

permission
 ▪ Unsupervised learning – concepts and usage
● Supervised ML: Linear Regression
o Linear relationship between two variables
o Measures of association: Pearson’s coefficient
o Types of correlation
o Best fit line – SSE, SSR and SST
o Standard error, MAE and RMSE
o R2 and adjusted R2
o Multiple linear regression
o Advantages & disadvantages of Linear regression
o How to treat independent categorical variable
o Case study
● Logistic regression
o Logistic function – Sigmoid curve
o Log loss
o Odds ratio
o Advantages & disadvantages of logistic regression
o Multiple dimensions
o Setting up threshold
o Confusion matrix – Accuracy, Precision, Recall, Specificity
o F1 score
o Case study

● K-NN classification
o Distance measures
o Euclidean, Manhattan, Minkowski distance measures
o Mahalanobis distance, Cosine similarity
o Radius neighbour & KD Tree classifier
o KNN for regression
o Advantages & Disadvantages of KNN
o Case study
o Optimizing the value of k
● Naïve Bayesian classifiers
o Conditional probability
o Independent events
o Baye’s theorem

Confidential – Not to be shared without

permission
 o Assumptions of Naïve Baye’s classifier
o Pros and cons of Naïve Baye’s theorem
o Case study
● SVM (Support Vector Machines)
o Hyperplanes
o Hard vs Soft margin
o Hinge loss
o Kernel tricks
o Margin and kernel
o Cover’s theorem
o Strength and weakness of SVM
o Case study

ENSEMBLE TECHNIQUES
● Decision trees
o Decision trees for classification and regression
o Entropy
o Information gain – Gini Index
o Overfitting decision trees
o Regularization methods – Pruning
o Regression trees
o Case study
● Ensemble methods
o Introduction to Ensemble methods
o Bagging
o Boosting techniques
o Ada boosting
o Case study

● Random Forest classifier

o Case study

UNSUPERVISED LEARNING
● K Means clustering
o Clustering

Confidential – Not to be shared without

permission
 o Distance measures – Manhattan, Euclidean, Chebyshev, Minkowski
o Types of clustering – Connectivity based & Centroid based
o Distance calculations
o Cluster definition
o Lloyd’s algorithm
o Elbow method
o Importance of scaling
o More distance measures – Mahalanobis, Jaccard distance
o Strength & Weakness of K Means clustering
o Visual analysis
o Silhouette score
o Dynamic clustering
o Case study

● Hierarchical clustering
o Connectivity based clustering
o Dendrogram
o Types of Linkages
o Cophenetic correlation
o Case study

● Principal component analysis

o Dimensionality reduction
o Scaling and Covariance matrix
o Eigen value and Eigen vectors
o Elbow method
o Feature elimination vs Feature extraction
o Case study

FEATURISATION, MODEL SELECTION & HYPER PARAMETER

TUNING
● Feature engineering
o Analytics based table
o Outlier treatment
● Model selection and tuning
o Variance vs Bias trade-off
o Up sampling and down sampling
o SMOTE

Confidential – Not to be shared without

permission
 o T Link
● Model performance measures
o Confusion matrix – Accuracy, specificity, recall & precision
o ROC curve and AUC curve
o F1 score
o Power of the test
o Hands-on – Confusion matrix
● Regularising Linear models
o L1 and L2 normalization
o Ridge and Lasso regression
o Hands-on – Ridge and Lasso

● ML pipeline
o Hands-on on ML pipeline
● Bootstrap sampling
● Grid search CV
● Randomized search CV
● K fold cross validation
o Implementing K fold cross validation
o Configuring K
o LOOCV
o Hands-on – K fold, LOOCV

RECOMMENDATION SYSTEMS
● Introduction to Recommendation systems and its applications
● Types of recommendation systems
● Popularity based recommendation system
● Market Basket analysis
o Concept of Market Basket analysis
o Support, Confidence and Lift
o Case study
o Precision, Recall & Accuracy
● Content based recommendation system
o Cold start problem
o Bag of Words approach
o TF - IDF
o Similarity measures – Jaccard, cosine similarity & Pearson’s coefficient
o Case study

Confidential – Not to be shared without

permission
● Collaborative filtering
o User – User collaborative filtering
o User – Item collaborative filtering
▪ Matrix factorization based approach using SVD
o Advantages and disadvantages of collaborative filtering
o Case study
o ROC curve
o Precision – Recall curve
● Hybrid recommendation systems

NEURAL NETWORKS BASICS

● Introduction to Perceptron & History of Neural networks
● Activation functions
o Sigmoid function
o ReLU
o Tanh
o Softmax
o Leaky ReLU
● Gradient descent
o Using Gradient descent to minimize loss function
o Vanilla, Stochastic and batch gradient descent
o Mini batching
● Learning rate and tuning
o Learning rate decay
o Momentum
o Sadle points, Hessians and local furrows
● Optimizer functions
o SGD optimizer
o Adam optimizer
● Introduction to Tensor flow
o Understanding computational graph
o Using Tensor flow to build a linear regression model
● Introduction to Keras
o Using Keras to build a simple Neural network
● Back propagation and chain rule
o Jacobian matrix
o Propagation of loss gradient
● Fully connected layer
● Softmax

Confidential – Not to be shared without

permission
● Cross – Entropy
● Weight initialization
o Problem with zero weights initialization
o Xavier initialization
o He initialization
● Regularization
o Batch Normalization
o Mini batching
o Drop outs
o Inverted drop outs
● Babysitting the neural network
● Hands-on demo for tuning the hyper parameter for Neural networks

COMPUTER VISION
● Introduction to Image data
o What is a Pixel?
o Saving & loading an image file
o Histogram
o Filters
● Introduction to Convolutional Neural Networks
o ANN vs CNN
o Translational Invariance
o Convolution operation
▪ 2D convolution
▪ Convolution on images
▪ Demonstration of forward and backpropagation in CNNs
▪ Dilated convolution
▪ Padding
▪ Dilated convolution
o Pooling
▪ Max pooling and average pooling
▪ Forward and backpropagation in pooling layers
o Hands-on demo of basic CNN
● Famous CNN architectures
o LeNet
o ImageNet challenge
o AlexNet - Architecture
o VGGNet - Architecture
o InceptionNet - Architecture

Confidential – Not to be shared without

permission
 o ResNet - Architecture
● Transfer Learning
● Case study – Transfer Learning on MNIST
● Object detection
o Object localization
o RCNN, Fast RCNN and Faster RCNN
o YOLO – Network architecture
o SSD (Single shot detection) – Network architecture
o Case study – Bounding box regressor
● Semantic segmentation
o What is Semantic segmentation
o Recap of CNN
o Deconvolution in detail
o UNet architecture
o Case study – One shot learning
● Other variants of convolution
o 1 * 1 convolution
o Concept of inception and separable convolution used in InceptionNet
o Concept of layer wise convolution used in MobileNet
● Metric Learning
o Concept and properties of metric learning
o Using CNNs as a feature extractor
● Siamese Networks
o Siamese Network architecture
o Distance measures
● Triplet Loss
o Why triplet loss is required
o Loss function for Siamese networks
o Using pre-trained models to build Siamese style networks
o Discussion on joint layers
o Case study – Arcadian Languages dataset

NLP
● Introduction to NLP
o Types of problems in NLP
o Challenges in NLP
o Workflow of dealing with NLP problems
o Common terminology used in NLP domain

Confidential – Not to be shared without

permission
● Preprocessing text data
o Tokenization
o Stop Words removal
o Stemming
o Lemmatization
o Word to features
● Bag of Words Model
● TF-IDF
● Introduction to Language models
● N-grams
o Markovian assumption
o Bi-grams and tri-grams
● Word2Vec
o Need for deep learning models
o CBOW (Continuous Bag of words)
o Skip gram
● GLOVE
o GLOVE as an alternative to Word2Vec
o GLOVE explanation and cost function
o Comparing the performance of GLOVE and Word2Vec
o Hands-on demo – Word embedding
● Case study - POS Tagging & Named Entity Recognition
● Introduction to Sequential models
o What is sequential data?
o Challenges of using traditional ML models to solve sequential data problems
● Need for memory in neural networks
● Types of sequential models – One to many, many to one, many to many
● Recurrent Neural networks (RNNs)
o RNN architecture
o Backpropagation through time (BPTT)
o Vanishing & Exploding gradients in RNNs
● Long Short Term Memory (LSTM)
o Concept of gates
o Forget gate
o Input/ output gate
o Cell state
o Forward propagation and backpropagation of LSTM
o Hands-on – POS tagging using LSTM
● GRU
● Applications of LSTMs
● Case study – Sentiment analysis using LSTM
● Time series analysis

Confidential – Not to be shared without

permission
 o Case study
● Case study - Neural Machine Translation
● Advanced LSTM architectures
o Multi-layer LSTM architecture
o Bi-directional LSTM architecture
o LSTM with attention mechanism

VISUALIZATION USING TENSORBOARD

● What is Tensor board?
● Visualizing graphs, weights, bias & gradients
● T-SNE
● Occlusion Experiment
● Saliency and Activation maps
● Neural Style transfer

REINFORCEMENT LEARNING
● ​REINFORCEMENT LEARNING
● What is reinforcement learning?
● Reinforcement learning framework
o Rewards
o Goals & actions
o State
o Policy
o Model
● Value-based methods - Q-learning
o Q learning – algorithm
o Q table
o Case study – Smart taxi
o Case study – Frozen lake
● Exploration vs Exploitation
● SARSA
● Q Learning vs SARSA

Confidential – Not to be shared without

permission
Introduction to GANs
● Introduction to GANs
● Auto encoders
● Deep Convolutional GANs
● How to train and common challenges in GANs
● Semi-supervised GANs
● Practical Application of GANs

SAMPLE LIST OF PROJECTS

1. A Campaign to sell Personal Loans

Identify potential customers for a personal loan product for a bank, allowing the
bank to design targeted marketing campaigns to increase conversion

2. Bank Note Analysis

The project is about building a classification model to predict the authenticity of
a bank note if it is counterfeit or genuine based on the characteristics of the
image of the bank notes.

3. Predict the onset of Parkinson’s disease

Parkinson’s Disease (PD) is a degenerative neurological disorder marked by
decreased dopamine levels in the brain. Traditional diagnosis of onset of
Parkinson’s disease is often difficult, as monitoring the onset requires repeated
clinic visits by the patient. The objective of the project is to build a Machine
Learning algorithm over a recording dataset to predict the onset of the disease

4. Silhouette analysis
The purpose of the case study is to build a model which can classify a given
silhouette as one of three different types of vehicle, using a set of features
extracted from the silhouette.

5. Network Intrusion detection

Confidential – Not to be shared without

permission
 Build a network intrusion detection system and improve the accuracy of your
prediction using a series of unsupervised and supervised methods.

6. Customer Sentiment from Amazon Reviews

Build a recommendation model that can determine and predict the sentiments
of the customer from text reviews on Amazon’s website

7. Street House View Numbers

Recognizing multi-digit numbers in photographs captured at street level is an
important component of modern-day map making. More broadly, recognizing
numbers in photographs is a problem of interest to the optical
character recognition community. The objective of this project is to build a
python code for image classification from scratch to understand the nuances of
building and training a model and further to understand the advantages of neural
networks

8. Face Recognition
Facial recognition is a biometric solution that measures unique characteristics
about one’s face. Given an image or a video capture of a scene with one or
more faces, the project is designed to detect and classify each face as one of
the persons whose identity is already known or as an unknown face.

9. Fake news detection

Fake news is increasingly becoming a threat to our society. It is typically
generated for increasing commercial interests—attract viewers and collect more
advertising revenue. In this project, you will build a classifier model which can
predict whether a piece of news is fake by using sequential models in Natural
Language Processing

Confidential – Not to be shared without

permission