|
|
Einrichtungen >> Fakultät Wirtschaftsinformatik / Angewandte Informatik >> Bereich Angewandte Informatik >>
|
Lehrstuhl für Erklärbares Maschinelles Lernen
| |
xAI-DL-M: Deep Learning -
- Dozent/in:
- Christian Ledig
- Angaben:
- Vorlesung, 2,00 SWS, ECTS: 6
- Termine:
- Di, 12:00 - 14:00, WE5/00.019
bis zum 31.1.2023
- Voraussetzungen / Organisatorisches:
- MSc AI, MSc WI, MSc CitH
Sign up: VC Course
- Inhalt:
- Vorkenntnisse / Prerequisites:
Good working knowledge of programming (e.g., in Python); Recommended completion of modules: Lernende System / Machine Learning [KogSys-ML-M], Einführung in die Künstliche Intelligenz / Introduction to AI [AI-KI-B], Mathematik für Informatik 2 (Lineare Algebra) [KTR-MfI-2], Algorithmen und Datenstrukturen [AI-AuD-B]
Description:
Deep Learning is a form of machine learning that learns hierarchical concepts and representations directly from data. Enabled by continuously growing dataset sizes, compute power and rapidly evolving open-source frameworks Deep Learning based AI systems continue to set the state of the art in many applications and industries. The course will provide an introduction to the most relevant techniques in the field of Deep Learning and a broad range of its applications.
The lecture will be held in English. The following is a selection of topics that will be addressed in the course:
- Relevant concepts in linear algebra, probability and information theory
- Deep feedforward networks
- Convolutional Neural Networks
- Regularization, Batch Normalization
- Optimization (Backpropagation, Stochastic Gradient Decent) and Cost Functions
- Classification (binary, multiclass, multilabel)
- Object Detection & Segmentation
- Generative Modelling
- Attention mechanisms & Transformer Networks
- Evaluation of ML approaches
Goals:
In this course students will learn/recap some fundamentals from mathematics and machine learning that are critical for the introduction of the concept of Deep Learning. Participants will learn about various foundational technical aspects including optimization and regularization strategies, cost functions and important network architectures such as Convolutional Networks. Students will further get an insight into more advanced concepts such as sequence modelling and generative modelling. Participants will further learn about representative architectures of important algorithm categories, e.g., classification, detection, segmentation, some of their concrete use cases and how to evaluate them.
The lecture is accompanied by exercises and assignments that will help participants develop practical, hands-on experience. In those exercises students will learn how to implement and evaluate Deep Learning algorithms using Python and its respective commonly used libraries.
- Empfohlene Literatur:
- Ian Goodfellow, Yoshua Bengio, and Aaron Courville: Deep Learning, MIT Press, 2016
- Zhang, Lipton, et al.: Dive into Deep Learning (https://d2l.ai/)
Further literature will be announced at the beginning of the course.
|
| |
xAI-Proj-M: Masterprojekt Erklärbares Maschinelles Lernen [xAI-Proj-M] -
- Dozentinnen/Dozenten:
- Ines Rieger, Christian Ledig
- Angaben:
- Übung, 4,00 SWS, ECTS: 6
- Termine:
- Do, 14:00 - 18:00, WE5/05.005
- Voraussetzungen / Organisatorisches:
- Interest and registration
If you have questions or want to express interest, please send an Email with name and matriculation number to ines.rieger@uni-bamberg.de. Registration via central VC course
- Inhalt:
- Topic: Deep Learning Life Cycle
Degree Program: M.Sc. AI, M.Sc. WI, M.Sc. ISoSySc, M.Sc. CitH (6 ECTS)
Requirements: Successfully passed the exam to KogSys-ML-M or AI-KI-B (Introduction to AI)
Beneficiaries: Knowledge in programming (Python), practical / hands-on knowledge in deep learning, scientific writing, LaTeX
Description The project provides the opportunity to work in small groups of 3 students in a hands-on fashion. The goal is to understand and implement the different steps to successfully train a deep learning model.
We will focus on the advantages and disadvantages of the design choices in data-preprocessing, model training, and model evaluation. You will gain theoretical knowledge about the design choices as well as practical knowledge by implementing these steps.
For the implementation, you are expected use Python and the deep learning framework PyTorch. Other libraries are free to choose.
At the end of the semester, you will present your results and hand in a technical project report.
The project builds on and adds practical experience to the knowledge from corresponding lectures and exercises in the area of machine learning.
Goals Students will familiarize themselves with a specific aspect of robust, explainable machine learning systems. Participants will learn to tackle a research-oriented question or problem independently, with little guidance. This will often involve the critical tasks: literature review, preparation and examination of datasets, implementation and comparison of prototypes, quantitative and qualitative evaluation of approaches. Within small groups, participants will learn to coordinate their project in a team and get comfortable with best practices of software development (e.g., testing, VCS).
Documentation and presentation of the project will help to develop both oral (presentation) and written (technical project report) communication skills in a scientific environment. In comparison to the Bachelor Project this Master Project is more ambitious in terms of complexity of selected topics as well as expectations with respect to deliverables and presentations.
Format
TBD
Expected workload & Grading
The workload of this module is expected to be roughly as follows:
- Attendance of project meetings / presentation: 35h
- Literature review and familiarization with topic (individual and within the team): 20h
- Implementation of selected algorithm / methodology: 70h
- Preparation of presentation: 15h
- Written documentation and report: 40h
The grade will be determined in equal parts based on the presentation and report. Attendance of the presentations is mandatory.
|
| |
xAI-Sem-B1: Bachelorseminar Erklärbares Maschinelles Lernen -
- Dozentinnen/Dozenten:
- Sebastian Dörrich, Christian Ledig
- Angaben:
- Seminar, 2,00 SWS, ECTS: 3
- Termine:
- Mi, 14:00 - 16:00, WE5/03.004
- Voraussetzungen / Organisatorisches:
- Interest and registration
If you have questions or want to express interest, please send an Email with name and matriculation number to sebastian.doerrich@uni-bamberg.de. Registration via central VC course
Requirements:
none
- Inhalt:
- Topic: Medical Imaging
Motivation: Medical imaging is used to support the diagnosis and treatment of diseases without the need for surgery or other intrusive measures.
Widely used imaging modalities enable the visualization of the interior of the human body, which consequently allows doctors to assess a patient's anatomy including bones, organs, tissue, and blood vessels through non-invasive means.
Hence, medical imaging can be used to detect a disease, help determine whether surgery is needed, locate tumors, find blood clots and other blockages, or assist doctors during interventions.
Acquired images can be collected and grouped together to create pathology-specific databases for the differentiation of abnormalities from normal anatomy, or the development of new procedures and approaches.
In a variety of research efforts, this often requires that patient data is made publicly available.
However, since this type of data is extremely sensitive, its storage and usage are thus restricted which in return poses a key challenge current researchers must face every day when working with medical data.
Topics can cover different aspects of the imaging process, including physics of image acquisitions (X-ray, MR, US, ...), reconstruction algorithms (e.g., backprojection), data storage formats (2D vs 3D, DICOM), modality specific benefits and challenges, etc. As such you can get answers to the following practical questons by attending the seminar:
What is medical imaging? How does Computed Tomography (CT) work?; When should we do an X-Ray scan rather than a CT or MRI?; When could Magnetic Resonance Imaging (MRI) be dangerous for patients?; What is the difference between MRI and fMRI?; Is ultrasound (US) suited for which type of tissue? When are PET and SPECT scanners used? What is an angiography system?; How can medical imaging support the detection and treatment of cancer?; What is Hybrid Imaging? Why do we need so many different imaging modalities?
Goals
In this seminar, you will learn about commonly used imaging modalities (e.g., MRI, CT, X-ray, US, …) used in healthcare by understanding their underlying physics, functionalities, and image acquisition processes.
You will further explore publicly available medical datasets of various anatomical regions while analyzing their different structures and formats.
In the end, you will be able to use your acquired knowledge to explore the potential as well as challenges of using medical data for current research. This seminar can be an essential building block if you are interested in building AI systems for healthcare applications.
Format
The presentations for this seminar will be conducted as block seminar. Dates TBD.
We will meet in the beginning of the semester to discuss work areas and assign concrete topics to each participant. You will be provided pointers to literature and then independently familiarize yourself with the assigned topic. Towards the end of the semester you will:
- present your topic as a 30 minute presentation and
- submit a written report of approximately 8 pages.
The presentations will be conducted as a block seminar towards the end of the semester. The weekly hours mentioned in the module description are an optional time slot to get support, guidance and feedback on your topic (as required).
Expected workload & Grading
The time (work load) of this module is expected to be roughly as follows:
- Attendance of seminar / presentation: 20h
- Literature review and familiarization with topic: 25h
- Preparation of presentation: 15h
- Written report: 30h
The grade will be determined in equal parts based on the presentation and report. Attendance of the presentations is mandatory.
|
| |
xAI-Sem-M1: Masterseminar Erklärbares Maschinelles Lernen [xAI-sem-M1] -
- Dozent/in:
- Christian Ledig
- Angaben:
- Seminar, 2 SWS, benoteter Schein, ECTS: 3
- Termine:
- Wir streben an, diese Veranstaltung in Präsenz durchzuführen. First meeting October 20, 4pm ct, WE5/05.003 // Second meeting October 24, 2pm ct, WE5/02.005
- Voraussetzungen / Organisatorisches:
- Interest and registration
If you have questions or want to express interest, please send an Email with name and matriculation number to christian.ledig@uni-bamberg.de. Registration via central VC course
Requirements:
completed course "Lernende System / Machine Learning" or "Einführung in die KI / Introduction into AI"
- Inhalt:
- This is a joint seminar between Prof. Kainz (FAU Erlangen-Nuremberg) and Prof. Ledig (University of Bamberg). The seminar will take place at Bamberg ERBA Campus and FAU Campus. Initial topic selection will take place in a hybrid format in Bamberg/Erlangen (in person on each site). Final topic presentations will take place in two sessions, one in person in Bamberg, one in person in Erlangen.
Topic: Human-in-the-Loop Machine Learning w/ focus on Healthcare
Motivation: Human-in-the-Loop Machine Learning describes processes in which humans and Machine Learning algorithms interact to solve one or more of the following: Making Machine Learning more accurate Getting Machine Learning to the desired accuracy faster Making humans more accurate Making humans more efficient Aim of this seminar is to give students insights about state-of-the-art Active Learning and interactive data analysis methods. Students will independently explore specific topics, which are then presented and discussed in class. Several potential topics will be provided but students are also encouraged to propose their own topics (after discussion with course lead).
Topics covered will include but are not limited to:
Introduction to Human-in-the-Loop Machine Learning: Active Learning Strategies, Uncertainty Sampling, Diversity Sampling, Other Strategies
Annotating Data for Machine Learning: Who are the right people to annotate your data?, Quality control for data annotation, User interfaces for data annotation
Transfer Learning and Pre-Trained Models: What are Embeddings?, What is Transfer Learning?
Adaptive Learning: Machine-Learning for aiding human annotation, Advanced Human-in-the-Loop Machine Learning
Goals
In-depth knowledge of aspects of human-in-the-loop machine learning, including deeper insight into current research. A capability to work independently on application-driven projects. To use a holistic view to critically, independently and creatively identify, formulate and deal with complex issues. To create, analyse and critically evaluate different technical/architectural solutions. To integrate knowledge critically and systematically. To clearly present and discuss the conclusions as well as the knowledge and arguments that form the basis for these findings in written and spoken English. A consciousness of the ethical aspects of research and development work.
The focus of the seminar will be biased towards approaches based on computer vision algorithms and medical image processing.
Format
The presentations for this seminar will be conducted as block seminar. Dates of final presentationsTBD.
We will meet in the beginning of the semester to discuss possible work areas and assign concrete topics to each participant. You will be provided pointers to literature and then independently familiarize yourself with the assigned topic. You will:
- present your topic as a 20 minute presentation (+5 min questions) and
- submit a written report of approximately 8 pages.
- The goal is to run the seminar in English including presentations and the written report.
The presentations will be conducted as a block seminar towards the end of the semester.
The weekly hours mentioned in the module description are an optional time slot to get support, guidance and feedback on your topic (as required).
Expected workload & Grading
The time (work load) of this module is expected to be roughly as follows:
- Attendance of seminar / presentation: 20h
- Literature review and familiarization with topic: 25h
- Preparation of presentation: 15h
- Written report: 30h
The grade will be determined in equal parts based on the presentation and report. Attendance of the presentations is mandatory.
|
|
 |
|
|
UnivIS ist ein Produkt der Config eG, Buckenhof |
|
|