Machine Learning is reshaping how we live. You may not know it but machine learning is all around you.

The application for example, In the case of Game of Thrones(GoT), we are given the storyline of GoT until now, we want to predict the likelihood of death of characters in GoT. How we can solve this problem?

Predicting the death of characters

From data to numerical attributes
Feed attributes into the computer
Let the computer predict the future

From Data to Numerical Attributes

The idea is we define character attributes that (potentially) might be relevant for prediction of a character’s death, for example, the Age, House Stark, House Targaryen, Married status, How popular in GOT wiki, etc.

Each Character is now a vector of attributes, for example

| House Stark |     |  1  |
| Age         |     |  36 |
| Married     |  =  |  1  |
| Female      |     |  0  |
| ...         |     | ... |

Then the color encoding, dead or alive,

Feed attributes into the computer

The second step is the computer now computes a smooth separation of the past data. between the death dan alive.

Let the computer predict the future

On the third step, for a character that wants to predict the likeliness of death, we use the separation computed by the computer.

What is Machine Learning?

Learning is the act of acquiring new knowledge, behaviors, skill, values, or preferences and may involve synthesizing different types of information. Or from the other sources, Learning is the ability to learn is possessed by humans, animals, and some machines.

So, what is machine learning, Machine learning is a field of study that gives computers the ability to learn without being explicitly programmed (Arthur Samuel). Computational methods using experience to make accurate predictions (Mhryar Mohri et al). Set of techniques that allow a computer to acquire or improve its ability to perform a task by automatically extracting knowledge from data.

Example of Machine Learning problems

Robot Learning Data: Sensor data gained by robots roaming their environments Goal: Robots learning to better navigate
Data Mining of electronic health records Data: Electronic health records of patients Goal: Learn to predict which therapies work best for which diseases
Speech Recognition Data: annotated recordings of speech (smartphone, customer service hotlines, etc) Goal: Better understand your speech based on experience listening to you
Optical Character Recognition Data: human-annotated handwritten digits or letters Goal: better recognize your handwriting

and many more example

Machine learning has a strong impact on all kinds of applications

What can Machine Learning be for you

Core tasks in ML: Theory > Algorithms > Applications

Theory, Analyze learning algorithms using techniques from probability and statistical learning theory, in order to understand them better
New Algorithms, Design of algorithms that learn faster or more accurately
Applications, Get learning algorithms working in applications

Machine learning is interdisciplinary (algorithms, AI, statistics, numerical mathematics, all kinds of application, etc.)

Basic Terminology

Data = Input + Labels, The data consists of inputs and labels.

Input, the raw data instances x₁, …, x_n (e.g. source code of computer programs)
Labels, their annotations y₁, …, y_n (malware: yes or no)

Training and prediction

Training, use all data (input and labels) to train the computer

Prediction, use the trained computer to predict the (unknown) labels for new inputs.

Formal Problem Setting

Let X and Y be some sets (called input space and label space, respectively)

Training data = \[ { (x_1, y_1), …,(x_n, y_n) } \in \mathbb R^d\]

The x₁, …, x_n is called inputs and y₁, …, y_n is called labels.

The goal of machine learning is to write a computer program that learns from the training data a function that accurately predicts on future or unseen data (data test)

the function is called the prediction function, or simply classifier, if Y is a finite set, the elements in Y are then called classes.

The computer program is synonymously also called learning machine or learning algorithm

How can we formally capture whether or not the computer predicts well

Let f be the classifier output after training the computer using the training data. For a new input x with label y the classifier f error when f(x) ≠ y. The error probability P[f(x) ≠ y] measures the quality of the classifier f.

We would like to design computer programs that compute f with error probability P[f(x) ≠ y] as small as possible!

Evaluation and Estimator for Error probability

t-times k-fold cross validation (cv)

function CV(t, k, training_data) 
  for i = 1:t do
  //Randomly split the samples into k sets of the same size (“folds”) 
    for j = 1:k do
    //Use jth fold as test set and union of all remaining folds as training set. Train classifier on training set and predict on test set
    end for
  end for

  return average classification accuracy and standard deviation 
  //(over the k · t many runs)
end function

Conclusion

Learning from experience
Machine Learning = making computers learn from data (usually to make accurate predictions of the future)
Example: k-nearest neighbor algorithm
Measure the quality of algorithms using cross-validation
ML: Strong impact on all kinds of applications
Many opportunities in this interdisciplinary field (algorithms, theory, applications,…)
Machine learning plays a key role in technology

Machine Learning