Introduction to the course: Table of Contents

Impact of AI

Challenges and opportunities

• Usually called the next electricity.
• AI is changing all areas of the society: communications, economy, politics.
• Very strong impact in learning, teaching, and education.
• Profound impacts on future labour markets and competence requirements.
• In 2016, more than 1600 startups specialized in IA were registered in the word.
• Race for supremacy in the word. Among the advanced countries: EUA, Canada, China, UK, Israel.

What is machine learning?

Definitions

Machine learning

Classification criteria for ML Methods

• Whether or not they are trained with human supervision (supervised, unsupervised, semi-supervised and reinforcement learning).
• Whether or not they can learn incrementally on the fly (online vs batch learning).
• Whether they work by simply comparing new data points to known data points, or instead they detect patterns in the training data and build a predictive model, much like scientists do (instance based vs model based learning).

Supervised learning

Renting a shared flat in Donosti

• After four years studying in San Sebastian a universitary student, not original from Donosti, perhaps from Ghana, has to find a flat again .
• He has lived in (8 flats already) so he knows about "good" and "bad" choices for a flat.
• Some memories are great (e.g., cool people he has lived with) and some are awful (the no-so-cool-flatmate's cat that used to eat his dinner).
• Therefore, he wants to use all his experience to predict, for three candidate flats, which ones could be a good choice.
• He can pose this flat selection question as a supervised classification problem.

Types of Machine Learning

Predicted or supervised learning

1. The goal is to learn a mapping from imputs \(x\) to ouputs \(y\), given a labeled set of input-ouput pairs \(D = {(x^i,y^i)}_{i=1}^{N}\). \(D\) is called the training set \(N\) is the number of training examples.
2. Usually, each \(x^i\) corresponds to a vector of numbers called features , attributes , or covariates.
3. It is assumed that the response variable or target variable \(y^i\) is a categorical or nominal variable from some finite set \(y^i \in \{1,\dots,C \}\) or that \(y^i \in \mathcal{R}\) .
4. When \(y^i\) is categorical the problem is known as classification. When, it is real-valued, the problem is known as regression.

Types of Machine Learning

Descriptive or unsupervised learning

1. We are given a set of inputs \(D = {(x^i)}_{i=1}^{N}\) and the goal is to find interesting or relevant patterns in the data.
2. This is a less well-defined problem usually called knowledge discovery.
3. Learning is done without any feedback or supervision from the user.
4. The most common unsupervised learning task is clustering.

Types of Machine Learning

Unsupervised learning problems

1. Clustering: To group points in the dataset according some criterion or similarity measure.
2. Biclustering: Do clustering taking into consideration two different criteria.
3. Dimensionality reduction or latent-factor discovery: Find an alternative representation of the problem in lower-dimensional space keeping as much information as possible.
4. Outlier detection: Detect observations that deviate markedly from other members of the sample in which it occurs.
5. Event or change-point detection: For time series, the identification of time points at which a behavior change occurs.
6. Discovering graph structure.

Types of Machine Learning

Reinforcement learning

1. An agent is situated in a given environment and has to make a sequence of decisions (or actions). For each action it receives an observation and a reward (or punishment) from the environment.
2. The goal of RL techniques is to train the agent so to maximize the cumulative reward to receive from the environment.
3. The environment is usually formalized as a partially observable Markov decision process (POMDP).
4. RL is deeply rooted in psychological and neuroscientific perspectives on animal behaviour, of how agents may optimize their control of an environment .

Types of Machine Learning

Other RL applications

1. Robotics.
2. Control (e.g., helicopter control).
3. Vehicle routing.
4. Human Computer Interaction (e.g., dialogue management).
5. Economics (e.g., trading simulations).
6. NN architecture search by network transformation.
7. Targeted marketing.

Types of Machine Learning

Genetic programming

1. GP is a method to create programs that should fullfill a prespecified objective.
2. A variety of program representations can be used, e.g., trees, program grammars, etc.
3. It can be used for solving classification and regression problems.
4. A measure of the program quality (fitness ) is defined and the search in the space of programs is posed as an optimization problem .
5. Genetic operators (e.g., mutation, crossover, and mutation) are used to search for the optimal program.
6. GP can be considered an evolutionary algorithm .

Supervised learning

Examples of GP algorithms

1. Tree-based genetic programming
2. Strongly typed genetic programming
3. Grammatical evolution
4. Probabilistic Incremental Program Evolution (PIPE)
5. Stack-based genetic programming
6. Cartesian genetic programming

Course objectives

Objectives

1. Enable the student to understand, develop, and implement models and algorithms capable of autonomously learning.
2. Present the main paradigms of machine learning approaches and the classes of problems where they can be applied.
3. Teach machine learning methods based on neural networks.
4. Introduce the most relevant types of neural networks, explaining the rationale behind their conception and scope of application.
5. Introduce and show how to implement deep neural networks.
6. Cover real-world applications of deep neural networks and hot topics in this area.

Course organization

Main modules

1. Introduction to machine learning.
2. Neural networks background (Biological and mathematical foundations).
3. Introduction to Neural Networks.
4. Neural networks paradigms.
5. Deep learning with neural networks.
6. Real-world application of deep neural networks.
   • Invited lecture by an AI expert most likely from the field of Natural Language Processing (NLP)
7. Advanced topics in neural networks studies.
8. Consolidation (Oral presentations and discussion of projects).

Course organization

Resources

1. The Machine learning and Neural Networks course has a virtual space in Egela, where links to useful software, bibliography, and problem repositories will be available.
2. After each class the slides will be available from the virtual space of the course.
3. Practical classes will be based on the Jupyter Notebooks framework with examples and exercises implemented in Python.
4. Jupyter notebooks will also be available from the virtual class after each practical class.
5. Students are encouraged to use the forum of the virtual class to exchange information, and ask and answer questions about the classes.
6. Every exchange during the course (e.g., questions, emails, oral presentations, use of the forum, etc.) should be in English.
7. Students are advised to create an account in github to develop, document, and keep up to date repositories containing the course projects that require an implementation.

Project characteristics

• Each of the three projects should correspond to a different topic . Topics are: 1) Machine learning. 2) Neural Networks. 3)Deep NNs.
• An ordinary project includes:
  1. A brief description of the project written in latex (between 4 and 12 pages) with latex style files available from Egela.
  2. The Python code (available from github or submitted to Egela in .zip format).
  3. A commented Jupyter notebook showing how the code solves the problem.
• Students can select a project of each topic from the available list.
• Each project will be available to a maximum of only two teams.
• An example of project submission (latex document, pdf, and commented notebook) will be available from Egela.

Project evaluation

Individual and team projects

• Projects can be individual or presented by a team of two students. In very exceptional cases (Ask me), 3 students.
• The same team can not present more than two projects.

Marks

• Minimum of 3 (correct) projects to pass the continuous evaluation course.
• Projects are assigned different difficulty levels: 1) Project of normal difficulty. 2) Medium degree of difficulty. 3) Very difficult project.
• The project difficulty will be taken into account for the final mark.
• For increasing the chance of a higher mark \( ( \geq 8)\), students are advised to include among the three selected projects, some of medium and high difficulty and/or to submit for evaluation 4 or 5 projects.