1 - Artificial Intelligence Foundations: Machine Learning

(A) What is Machine Learning?

Question 1 :- Drithi has formal badminton lessons every Tuesday afternoon, and plays with
whoever is available on Saturdays. Which type of learning is she using?

ANS :- semi-supervised

Question 2 :- In the learning process, what would you typically do after implementing a rule?

ANS :- Observe feedback.

Question 3 :- Which sort of problem is most explicitly amenable to a solution by machine

learning?

ANS :- analyzing menu selection preferences

Question 4 :- Instead of explicit, stepwise instructions, what are the fundamental inputs to a
machine-learning process?

ANS :- data

(B) Different ways of Machine Learning?

Question 1 :- What distinguishes supervised machine learning from other types of machine
learning?

ANS :- using labeled data for training

Question 2 :- How are induction and transduction different?

ANS :- Transduction uses more information and produces more specific rules

Question 3 :- A company has a variety of efforts that it uses to find new antiviral medications.
Which effort best exemplifies unsupervised learning?

ANS :- the screening of a large collection of botanical extracts

Question 4 :- Why would you choose to not use semi-supervised learning all the time?

ANS :- It can lead to larger errors or confusion

Question 5 :- What does Q-learning use instead of training or testing data?

ANS :- simulation or experience

 (C) Popular Machine Learning Algorithms

Question 1 :- Why is k-nearest neighbor also called lazy learning?

ANS :- It uses a lot of computation for every instance.

Question 2 :- What characterizes an outlier in cluster analysis?

ANS :- not being close to any centroid

Question 3 :- Why are dog weight and height not necessarily good choices for predictors in
naïve Bayes methods?

ANS :- They are correlated

Question 4 :- Why are regression methods not considered good examples of machine
learning?

ANS :- They are based on statistical predictions

Question 5 :- What should you do if your decision tree has too much entropy?

ANS :- Add or substitute predictors

Question 6 :- Which one of the following do you use when you look for trends instead of
trying to classify data into different groups?

ANS :- regression problems

Question 7 :- Salim wants to predict the best harvest date for his orchard based on prior
weather reports and harvest histories. Which type of tool does he require?

ANS :- regression analysis

Question 8 :- Which of the following is instance-based (or lazy learning)?

ANS :- K-Means Clustering and K-NN

(D) Applying Algorithms

Question 1 :- Which scenario is an example of high bias and low variance?

ANS :- predictions that are almost always 7.5 pounds too high
Question 2 :- Hugo's modeling strategy uses k-nearest neighbor, followed by regression
analysis. What does his strategy exemplify?

ANS :- stacking or boosting

Question 3 :- Lydia has produced a model with 100 predictors that is capable of almost
perfectly fitting the 200 observations in the training data. How would you improve Lydia's
model?

ANS :- by reducing its complexity

(E) Common Challenges

Question 1 :- Most of Atul's experience is with decision trees and regression. Which strategy
should he embrace as manager of a new project?

ANS :- applying decision trees and regression to the new problem

Question 2 :- You finished fine-tuning your model with training data, and are eager to show
the results to your business team. What should you do instead?

ANS :- Show predictions for the testing data

Exam :-

Question 1 :- In _____, you are using _____ to help your program find patterns in _____.

ANS :- machine learning; artificial intelligence; massive datasets

Question 2 :- What does the training data, which is a smaller chunk than the test data, help
you find?

ANS :- patterns

Question 3 :- If you are training the machine to predict how long it takes you to drive home,
which one of the following is the dependent variable - weather, time of day, holiday, or
commute time?

ANS :- commute time

Question 4 :- What is being optimized in Q-learning?

ANS :- the quality of the outcome or performance

Question 5 :- With what type of learning will you need to have access to a lot of data?

ANS :- unsupervised

Question 6 :- What do you call the set environments in Q-learning?

ANS :- states

Question 7 :- Joaquim needs a photograph analysis tool to identify the plant species
corresponding to individual seeds. Which type of problem does he have?

ANS :- multiclass classification

Question 8 :- Which one of the following statements is true?

ANS :- all of these answers

Question 9 :- Decision trees can be used for _____ classification challenges with _____ machine
learning.

ANS :- binary; supervised

Question 10 :- Which of the following leads to being consistently wrong in an inconsistent

way?

ANS :- high bias / high variance

Question 11 :- What statement is correct?

ANS :- Naïve Bayes is based on conditional probability

Question 12 :- How do k-means clustering methods differ from k-nearest neighbor

methods?

ANS :- Clustering methods are unsupervised.

Question 13 :- Which of the following is the best fit when you have lots of unlabeled data?

ANS :- K-means clustering

Question 14 :- What is the total score for a hound corresponding to the following weights
and individual scores?

Multiplier Hound

Hair 3.5 0.8

Height 2.0 0.6

Weight 1.5 0.4

ANS :- 4.6
 2 - Applied Machine Learning: Algorithms
(A) Review of Foundations

Question 1 :- A concrete model is created when an algorithm is fit to data.

ANS :- TRUE

Question 2 :- Which is NOT a reason why we had to determine whether Age was missing at
random?

ANS :- The model will treat it different if it's missing at random

Question 3 :- How could we tell Cabin was NOT missing at random?

ANS :- splitting power on survival rate

Question 4 :- Scikit-learn's train_test_split method is capable of splitting a single dataset into

three data subsets.

ANS :- FALSE

(B) Logistic Regression

Question 1 :- Either Linear Regression or Logistic Regression are reasonable choices to

model a binary target variable.

ANS :- FALSE

Question 2 :- Logistic Regression is a good choice when you have a massive amount of data
or you're trying to solve a state of the art problem.

ANS :- FALSE

Question 3 :- What does NOT happen with a very low value of C?

ANS :- more likely to memorize examples in training set

Question 4 :- How many individual Logistic Regression models were fit using GridSearchCV?

ANS :- 35
 (C) Support Vector Machines

Question 1 :- The kernel trick is used every time you fit SVM to data.

ANS :- FALSE

Question 2 :- SVM's main competitive advantage is its performance on short-and-fat data or

when there are a lot of outliers.

ANS :- TRUE

Question 3 :- Which of the following does NOT happen with a very low value of C?

ANS :- more likely to overfit

Question 4 :- Our Titanic data appears to be linearly separable.

ANS :- TRUE

(D) Multi-layer Perceptron

Question 1 :- Which is NOT a type of layer in a Multilayer Perceptron?

ANS :- Regularization layer

Question 2 :- Which is NOT a benefit of Multilayer Perceptrons?

ANS :- It's fast to train.

Question 3 :- Which is NOT a side-effect of adding more hidden layers/nodes?

ANS :- worse training accuracy

Question 4 :- The activation function controls how many hidden layer nodes are used in
training.

ANS :- FALSE

(E) Random Forest

Question 1 :- What is the main reason you sample both examples AND features for each
decision tree in a Random Forest?

ANS :- encourage independence of trees

Question 2 :- Random Forest is a tremendously flexible model that makes for a really good
initial benchmark model.

ANS :- TRUE

Question 3 :- Increasing n_estimators or max_depth will ALWAYS decrease the training error.

ANS :- TRUE

Question 4 :- If I fit two Random Forest models with the exact same hyperparameter settings
on the exact same data, I should get exactly the same model and performance.

ANS :- FALSE

(F) Boosting

Question 1 :- The trees built in Gradient Boosted Trees can be fit in parallel.

ANS :- FALSE

Question 2 :- Gradient Boosted Trees' ability to learn from its own mistakes also drives it to
overfit to outliers.

ANS :- TRUE

Question 3 :- A small learning rate will ensure you always find the optimal model.

ANS :- FALSE

Question 4 :- Which statement is true regarding Random Forest and Gradient Boosted Trees?

ANS :- Random Forest does better with few, deep trees, and Gradient Boosted Trees does
better with many, shallow trees.

(G) Summary

Question 1 :- What is the No Free Lunch Theorem?

ANS :- No algorithm works best for every problem.

Question 2 :- Which statement is NOT true?

ANS :- Gradient Boosted Trees takes a long time to make predictions.

Question 3 :- Recall and Precision are only necessary if you don't have access to Accuracy.

ANS :- FALSE

3 - Applied Machine Learning: Foundations

(A) Machine Learning Basics

Question 1 :- What is one of the key purposes of machine learning?

ANS :- generalize to unseen examples

Question 2 :- Which is NOT a reason why we use Python for machine learning?

ANS :- It's the fastest language available.

Question 3 :- Which is NOT accurate?

ANS :- Strong AI is the equivalent of machine learning.

(B) Exploratory Data Analysis and Data Cleaning

Question 1 :- Exploratory data analysis is performed only to inform data cleaning.

ANS :- FALSE

Question 2 :- Which is NOT a reason why we had to determine whether Age was missing at
random?

ANS :- The model will treat it different if it's missing at random.

Question 3 :- Looking at mean values overstates the impact of Fare on whether somebody
survived or not.

ANS :- TRUE

Question 4 :- If we do NOT include "inplace=True" in our drop statement, then we need to

assign the alteration to a new dataframe.

ANS :- TRUE

Question 5 :- How could we tell Cabin was NOT missing at random?

ANS :- splitting power on survival rate

Question 6 :- Embarkation is a causal factor in determining whether somebody survived.

ANS :- FALSE

Question 7 :- Which is NOT a required argument for the where method from numpy?

ANS :- inplace=True (or False)

(C) Measuring Success

Question 1 :- Model performance on the training set matters more than performance on the
validation set or test set.

ANS :- FALSE

Question 2 :- Scikit-learn's train_test_split method is capable of splitting a single dataset into

three data subsets.

ANS :- FALSE

Question 3 :- How many individual models will be built in standard 10-fold Cross-Validation?

ANS :- 10

Question 4 :- Which is NOT a commonly used performance metric for classification

problems?

ANS :- cross-validation

(D) Optimizing a Model

Question 1 :- The goal of model optimization is to tune model complexity to minimize total
error by reducing variance and bias.

ANS :- TRUE

Question 2 :- Which is NOT a characteristic of underfitting?

ANS :- memorizing training examples

Question 3 :- Which is NOT a characteristic of overfitting?

ANS :- not capturing underlying trend in data

Question 4 :- If test error is high, what is the best tool to determine whether you're
underfitting or overfitting?

ANS :- training error

Question 5 :- Which is NOT true of a hyperparameter?

ANS :- It is learned from data.

Question 6 :- Which is NOT true of regularization when applied appropriately?

ANS :- improves models performance on training data

(E) End-to-End Pipeline

Question 1 :- We fill missing values in the Age column with the average to ensure the model
isn't determining whether somebody survived or not based on Age alone.

ANS :- TRUE

Question 2 :- Keeping the Embarked feature in our data likely would have generated better
results on unseen data.

ANS :- FALSE

Question 3 :- Index=False should still be used even if you have a meaningful index column.

ANS :- FALSE

Question 4 :- Which of the following is NOT a benefit of Cross-Validation?

ANS :- faster than just fitting on the training set and evaluating on a validation set

Question 5 :- How many individual models are built using 5-fold Cross-Validation and 12
total hyperparameter combinations?

ANS :- 60

Question 6 :- If the best model based on Cross-Validation does NOT perform best on the
validation set, then you did something wrong.

ANS :- FALSE

Question 7 :- What is the most likely reason for test set performance to deviate in a
significant way from the validation set performance?

ANS :- limited data and/or data wasn't split randomly

 4 - Machine Learning and AI Foundations: Classification
Modeling
Introduction

Question 1 :- What is a gateway course?

ANS :- It is a gateway to all kinds of machine learning topics and broad data science.

Question 2 :- What are the three types of variables?

ANS :- nominal, ordinal, and scale variables

(A) The Big Picture : Defining Your Classification Strategy

Question 1 :- Which document is used to evaluate the performance of your models?

ANS :- Cross-Industry Standard Process for Data Mining

Question 2 :- In Tom Khabaza's 9 Laws of Data Mining, what is the fourth law?

ANS :- The fourth law states, "the right model for a given application can only be
discovered by experiment."

Question 3 :- What do machine learning and AI-based solutions need?

ANS :- They need accurate, well-chosen algorithms to perform classification correctly.

Question 4 :- Why should you use two categories?

ANS :- Two categories require just one linear formula.

Question 5 :- What is logistic regression?

ANS :- Logistic regression is where all of the inputs can be ranked from most to least
important.

Question 6 :- Why do binary classifications play an important role?

ANS :- because they are used in every project

 (B) How do I choose a “Winner”?

Question 1 :- How do you partition historical data?

ANS :- You divide the data or partition the data into 50% training data and 50% testing
data.

Question 2 :- Why should you use a lift chart?

ANS :- Lift charts assess models and all kinds of variations on a theme.

Question 3 :- What is the confusion matrix?

ANS :- The model makes errors. It's a technique for accessing the model's performance.

(C) Algorithms on Parade

Question 1 :- What is Stepwise Discriminant Analysis?

ANS :- It selects useful subsets of variables to evaluate the order of importance of

variables.

Question 2 :- Who uses logistic regression?

ANS :- Epidemiologists and Health Researchers

Question 3 :- What does K-Nearest Neighbor (KNN) do?

ANS :- KNN memorizes the location of all the cases.

Question 4 :- Which applications use deep learning?

ANS :- visual recognition and speech recognition

Question 5 :- Decision Trees are also called _____.

ANS :- greedy algorithms

Question 6 :- There are three common problems: interactions, missing data, and overfitting.
What is overfitting?

ANS :- Overfitting is when we want our models to be just complex enough.

Question 7 :- When using heterogeneous ensemble, which three models should you use?

ANS :- logistic regression, neural networks, and support vector machine

Question 8 :- What is Discriminant Analysis?

ANS :- You analyze the data when the criterion or the dependent variable is categorical,
and the independent variable is an interval in nature.

Question 9 :- Neural Networks use multilayer perceptrons. What are multilayer perceptrons?

ANS :- They consist of three layers of nodes: an input layer, a hidden layer, and an output
layer.

Question 10 :- What is Bayes theorem?

ANS :- Bayes theorem is the probability of an event, based on prior knowledge of

conditions that might be related to the event.

(D) Common Modeling Challenges

Question 1 :- What is feature selection when we are describing data reduction?

ANS :- It is the removal of poor and redundant predictors before modelling.