Machine Learning

Lecture 1: Introduction

Instructor: Dr. Farhad Pourkamali Anaraki

 Machine Learning, Pourkamali

What do you learn from this course?

• Concepts, tools, and intuition you need to implement programs capable of
learning from data

• From simple techniques such as linear regression to some of deep learning

techniques

• A hands-on approach with just a little bit of theory to use production-ready

Python frameworks

Public perception of ML Reality

2
 Machine Learning, Pourkamali

Course prerequisites
• We need some level of calculus and linear algebra
• Vector, matrix, inner product, matrix-matrix multiplication
• Derivatives of polynomials and exponential functions
• Chain rule for nding the derivative of a composite function

• Basics of probability and statistics

• Probability distributions (normal distribution, Bernoulli distribution, etc.)
• Mean, standard deviation, …

3
fi
 Machine Learning, Pourkamali

Anaconda
• https://www.anaconda.com/

4
 Machine Learning, Pourkamali

Google Colab
• Google Colaboratory, or “Colab” for short, allows you to write and execute
Python in your browser

• Zero con guration required

• Free access to GPUs and other computing resources
• Use terminal commands on Google Colab

• https://drive.google.com (for the rst time, select “Connect more apps”)

5
fi
fi
 Machine Learning, Pourkamali

Visual Studi Code

• Better alternative for “real” machine learning projects
• https://code.visualstudio.com/

6
 Machine Learning, Pourkamali

Course structure
• Part 1: Fundamentals of Machine Learning (Scikit-Learn)
• Review of Python’s main scienti c libraries: NumPy, pandas, and Matplotlib
• Steps in a typical machine learning project
• Learning by tting a model to data and optimizing a cost function
• Challenges of using machine learning systems
• Part 2: Neural Networks and Deep Learning (TensorFlow 2.0, Keras, and PyTorch)
• What neural nets are and what they are good for
• Building and training feedforward neural nets (basics and implementation)
• Convolutional nets for deep computer vision and recurrent neural nets (mainly
fundamental concepts)

7
fi
fi
 Machine Learning, Pourkamali

A little bit history

• Alan Turing gave quite possibly the earliest public lecture on this topic in 1947:
“What we want is a machine that can learn from experience”

8
 Machine Learning, Pourkamali

Why should we use machine learning?

• Let’s say we want to design a spam ltering system
• One approach is to use traditional programming techniques
• Look for words or phrases such as: credit card, free, bank information, “C U”
• Write a detection algorithm for each of the patterns
• Test your program and improve it

9
fi
 Machine Learning, Pourkamali

Why use machine learning?

• A spam lter based on machine learning automatically learns which words and
phrases are good predictors

• Possible bene ts: simpler, easier to maintain, and more accurate

• The examples that the system uses are called the training set
• We should also think of a performance measure

10
fi
fi
 Machine Learning, Pourkamali

Beyond prediction
• Machine learning can help human learn and discover patterns from large
amounts of data that are not immediately apparent

• This is usually called data mining

11
 Machine Learning, Pourkamali

A few applications of machine learning

• Autonomous navigation systems

• Analyzing images of items on a product line
• Analyzing medical images
• Automatically agging offensive comments on social media platforms
• Forecasting revenue
• Creating a recommendation engine
• Speech recognition (like Siri and Alexa)
• And many more!

12
fl
 Machine Learning, Pourkamali

Types of machine learning

• ML systems can be classi ed according to the amount and type of supervision
1. Supervised learning: training set includes the desired solutions or labels

• Classi cation: labels are classes (e.g., spam or ham)

• Regression: labels are numeric values (e.g., price of car)

13
fi
fi
 Machine Learning, Pourkamali

Types of machine learning

2. Unsupervised learning: ML system tries to learn without a teacher

• For example, based on the shopping history, detect groups of similar visitors

• Another example is novelty detection: nd new instances that look different

14
fi
 Machine Learning, Pourkamali

Types of machine learning

3. Semisupervised learning: plenty of unlabeled instances and few labeled

• Why? Labeling data is costly and time-consuming

• Application: Google Photo
• Tools: Amazon SageMaker Ground Truth

4. Reinforcement learning: select and perform actions, and get rewards or

penalties in return

• Find the best strategy known as policy to get the most reward

15
 Machine Learning, Pourkamali

Another categorization
1. Batch learning or of ine learning: the system must be trained using all the
available data

• Drawbacks:

• Take a lot of time, computing resources, and storage space

• Not suitable for rapidly changing data

2. Incremental learning: train the system incrementally by feeding data instances

sequentially

16
fl
 Machine Learning, Pourkamali

Another categorization
1. Instance-based learning: the system uses a similarity measure to compare new
cases to the learned examples

6. Model-based learning: build a model of the learned examples and then use that
model to make predictions

17
 Machine Learning, Pourkamali

Model-based learning
• Does money make people happier?

• Let us model life satisfaction as a linear function of GDP

• A simple linear model

• Examples

18
 Machine Learning, Pourkamali

Training a linear model using Scikit-Learn

• A few lines of code

19
 Machine Learning, Pourkamali

Summary of steps

• Study the data (data preparation)

• Select a model or learning algorithm
• Train on the available data and assess performance
• Apply the model to make predictions on new cases

20
 Machine Learning, Pourkamali

Challenges of machine learning

• Insuf cient quantity of training data
• Non-representative and poor-quality training data

• Irrelevant features: coming up with a good set of features

21
fi
 Machine Learning, Pourkamali

Challenges of machine learning

• Over tting the training data
• When the model is too complex relative to the amount and noisiness of data

• How to alleviate this problem?

• Simplify the model
• Gather more training data
• Reduce noise and outliers (i.e., data preparation)

22
fi
 Machine Learning, Pourkamali

Example of reducing overfitting

• Use a subset of data and regularization
• Regularization: balance between tting the data perfectly and keeping the
model simple enough

• Amount of regularization is controlled by a hyperparameter (must be set prior

to learning and remains constant during training)

• Tuning hyperparameters is extremely important

23
fi
 Machine Learning, Pourkamali

Challenges of machine learning

• Under tting the training data
• When the model is too simple

• How to solve this problem?

• Select a more powerful model
• Reduce constraints on the model such as regularization

24
fi
 Machine Learning, Pourkamali

Testing and validation

• The only way to know how a model will generalize to new cases is to actually try
it out on new cases

• Split the available data into two sets: training set and testing set
• The error on the test set is called the generalization error or out-of-sample
error

• If training error is low and generalization error is high, then over tting occurs
• How to choose suitable models and their hyperparameters?

25
fi
Basics of Python
https://github.com/farhad-pourkamali/machine-learning
Review of linear algebra and NumPy
https://github.com/farhad-pourkamali/machine-learning
Part 1: https://youtu.be/xSO0qOyvevc
Part 2: https://youtu.be/SeBRHg9ZrSs
 Machine Learning, Pourkamali

Introduction
• NumPy offers mathematical functions, random number generators, linear algebra
routines, and more

• ‘’de-facto” standards of array computing

• 1D (vector) and 2D arrays (matrix)

a1
a11 … a1m
a2
a= ∈ ℝn A= ⋮ ⋮ ∈ ℝn×m
⋮ an1 anm
an

28
 Machine Learning, Pourkamali

Creating arrays

• Notice the difference between 1D and 2D arrays

29
 Machine Learning, Pourkamali

Creating arrays

30
 Machine Learning, Pourkamali

Creating random numbers

• Two popular choices are uniform distribution between 0 and 1, and a univariate
normal distribution

31
 Machine Learning, Pourkamali

Array data
• NumPy’s ndarrays are ef cient because their elements must have the same type
• Data types: https://numpy.org/doc/stable/user/basics.types.html

32
fi
 Machine Learning, Pourkamali

Array data
• NumPy’s ndarrays are ef cient because their elements must have the same type

33
fi
 Machine Learning, Pourkamali

Reshaping an array

34
 Machine Learning, Pourkamali

Reshaping an array
• The reshape function returns a new array object pointing at the same data.

35
 Machine Learning, Pourkamali

Conditional operators

36