Statistical Learning

Code:

M4125

Acronym:

M4125

Keywords
Classification	Keyword
CNAEF	Mathematics and statistics

Instance: 2020/2021 - 2S

Active?	Yes
Web Page:	https://cmup.fc.up.pt/cmup/statlearning
Responsible unit:	Department of Mathematics
Course/CS Responsible:	Computational Statistics and Data Analysis

Cycles of Study/Courses

Acronym	No. of Students	Study Plan	Curricular Years	Credits UCN	Credits ECTS	Contact hours	Total Time
E:ECAD	14	PE_Estatística Computacional e Análise de Dados	1	-	6	42	162

Teaching language

Suitable for English-speaking students

Objectives

In most scientific disciplines, hypotheses are assessed and validated by applying statistical tools to experimental or observational data.

Two distinct paradigms dominate the statistical scenario: the frequencist and bayesian approach, both treated in detail in other ucs in this cycle of studies.

The goals of statistical learning are understanding and predicting. It falls into several categories, including supervised, unsupervised, semi-supervised and reinforcement learning. The objective of this course is to treat each of these categories, application areas of each one, methodologies of realization through neural network architectures (deep) and computational implementation using tensorflow and keras.

Various types of neural network architectures will be used: multilayer (feedforward); convolutive, recursive, generative and others. Network optimization and regularization processes and problems of practical application in various scientific and technological areas will be addressed.

Objectives and competences:

1. Distinguish clearly the different learning categories mentioned above
2. Distinguish the different neural network architectures referred to
3. Implement these architectures with tensorflow and keras

4. Given a specific situation, choose the appropriate learning methodology and its computational implementation.

Learning outcomes and competences

Learning outcomes and competences:

1. Distinguish clearly the different learning categories mentioned above
2. Distinguish the different neural network architectures referred to
3. Implement these architectures with tensorflow and keras

4. Given a specific situation, choose the appropriate learning methodology and its computational implementation.

Working method

Presencial

Program

1. Learning and Inference Methods. Supervised, Unsupervised, and Reinforcement Learnings, Semi-supervised Learning and Active Learning


2. Model Selection. Cross-Validation. Robust Learning


3. Neural Networks as Universal Machines. Multilayer Perceptrons: Architecture and Error. Backpropagation.


4. Recurrent Neural Networks. Backpropagation for Temporal Learning. Networks for Modeling Dynamic Systems.


5. Deep Convolutional Neural Networks.


6. Representation Learning. Autoencoders. Generative Adversarial Neural Networks.


7. Reinforcement Learning. Actor–Critic Model. Model-Free and Model-Based Reinforcement Learning. Temporal-Difference Learning. Q-Learning

Mandatory literature

Ovidiu Calin; Deep learning architectures. ISBN: 978-3-030-36721-3

Comments from the literature

Written notes and recorded classes will be published to support the curricular unit, authored by the head of the UC.

Teaching methods and learning activities

Teaching methodologies include formal interactive classes, promoting the expository components with interpellation and dialogue with students. Whenever necessary, lectures involving students in a more active process are included, and case studies / articles are used, using tools such as books, articles and the internet, seeking to guide students in a structured way to understanding of subjects; practical laboratory classes, where students perform experiments, implement crucial algorithms in the area and / or use data analysis programs that allow them to develop transversal and integrated laboratory and computer skills in the area of statistical learning. The teaching methodologies mentioned above will be adjusted to allow students to integrate the objectives of the course.

Evaluation Type

Distributed evaluation without final exam

Assessment Components

designation	Weight (%)
Apresentação/discussão de um trabalho científico	50,00
Teste	50,00
Total:	100,00

Amount of time allocated to each course unit

designation	Time (hours)
Frequência das aulas	30,00
Apresentação/discussão de um trabalho científico	12,00
Total:	42,00

Eligibility for exams

To obtain frequency, students have to

1. present an individual work on a UC theme to be agreed with the responsible teacher


2. onduct a written test on uc syllabus


3. Have participation in at least 50% s of classes taught in synchronous regime (in person or remotely).

Calculation formula of final grade

The final classification is obtained with the arithmetic mean of the classifications obtained in the following components: (i). Presentation / discussion of a scientific paper, and (ii). written test.
Each of these components will be valued with a maximum of 20 values, each having a weight of 50% in the calculation of the final classification.

The student must obtain a minimum of 8 values (a total of 20 values) in each of the two components mentioned.

Classification improvement

The classification improvement will be done through a written exam.

Observations

The course of operation of the course is subject to the limitations imposed by FCUP according to the evolution of the pandemic COVID19. We do not expect a 100% face-to-face operation. It may be in a B-Learning system if the circumstances so require.