2023-01-17 ~ 2023-01-20 Machine Learning for Computer Vision
For CAS AML participants as elective module 6.
Reiter
About this training
Machine Learning for Computer Vision
Image processing or computer vision is one of the domains in which Machine Learning (ML) and Deep Learning (DL) methods in particular have been applied the most successfully. The specific nature of images as data (both natural such as photographs or scientific such as in microscopy) composed of locally correlated information at different scales is a perfect fit for DL methods which have dramatically surpassed classical image processing approaches in the past decade. Such methods are used for a wide array of tasks including for example image classification, segmentation, denoising and even generation, and find applications in virtually every scientific and business area.
In this course we will explore these methods in three steps. First we will explore both concepts and tools that are specific to images compared to tabular or textual data. We will for example learn how to create and deal with image dataset or understand what convolutional DL layers are. Second, we will build from scratch a simple convolutional neural network (CNN), train and test it and see step by step how to improve it. This will allow us to become familiar with image-specific topics such input/output dimensionalities, receptive fields, augmentation etc. Third, after this deep dive in the inner-workings of a network, we will learn how to use higher-level software packages that give an easy out-of-the-box access to complex network architectures that can be used pre-trained or fine-tuned for specific applications.
In their personal project, participants will try to solve a real-world image-based problem using ML or DL. They are free to choose the type of problem and their dataset but ideally would work using data from their own field of work.The practical parts of this course are based on PyTorch and not Tensorflow (unlike most other courses of the program). This is an opportunity for participants to learn about this other main deep learning framework that they might encouter in the professional or academic work. Note that the choice of DL framework used the realise the personal project is left to the participants.
Image processing or computer vision is one of the domains in which Machine Learning (ML) and Deep Learning (DL) methods in particular have been applied the most successfully. The specific nature of images as data (both natural such as photographs or scientific such as in microscopy) composed of locally correlated information at different scales is a perfect fit for DL methods which have dramatically surpassed classical image processing approaches in the past decade. Such methods are used for a wide array of tasks including for example image classification, segmentation, denoising and even generation, and find applications in virtually every scientific and business area.
In this course we will explore these methods in three steps. First we will explore both concepts and tools that are specific to images compared to tabular or textual data. We will for example learn how to create and deal with image dataset or understand what convolutional DL layers are. Second, we will build from scratch a simple convolutional neural network (CNN), train and test it and see step by step how to improve it. This will allow us to become familiar with image-specific topics such input/output dimensionalities, receptive fields, augmentation etc. Third, after this deep dive in the inner-workings of a network, we will learn how to use higher-level software packages that give an easy out-of-the-box access to complex network architectures that can be used pre-trained or fine-tuned for specific applications.
In their personal project, participants will try to solve a real-world image-based problem using ML or DL. They are free to choose the type of problem and their dataset but ideally would work using data from their own field of work.The practical parts of this course are based on PyTorch and not Tensorflow (unlike most other courses of the program). This is an opportunity for participants to learn about this other main deep learning framework that they might encouter in the professional or academic work. Note that the choice of DL framework used the realise the personal project is left to the participants.
The practical parts of this course are based on PyTorch and not Tensorflow (unlike most other courses of the program). This is an opportunity for participants to learn about this other main deep learning framework that they might encouter in the professional or academic work. Note that the choice of DL framework used the realise the personal project is left to the participants.
Goal
- The goal of this course if to familiarize participants who already have some knowledge in machine learning with the specificities of the image processing domain, and give them tools to solve problems they might encounter in their area.
Learning objectives
After the course, participants can
After the course, participants can
- determine what type of approach can be used to solve their image-based problem
- conceptually understand the structure of classic neural networks
- use pre-made networks available in popular libraries such as fastai or keras as applied tools to solve problems.
Target group
- CAS Advanced Machine Learning participants
- Other interested people
Prerequisites
- Basic familiarity with Python and Jupyter notebooks
- Basic machine learning and neural network skills
- Own laptop
Methods
- Theoretical introductions
- Hands-on tutorials with Jupyter notebooks
- Project work with presentation
Format
- About 15-20 hrs presence (online possible)
- 30 hours project work
- Assessment as oral presentation of project work
Certificate
- A certificate will be delivered to participants who have attended the whole training
Coaches
- The coaches are local or external experts
Schedule
Tuesday
09:00 - 12:30
Wednesday
09:00 - 12:30
Thursday
09:00 - 12:30
Friday
09:00 - 12:30
09:00 - 12:30
Wednesday
09:00 - 12:30
Thursday
09:00 - 12:30
Friday
09:00 - 12:30
Practical information (time, location ...)
Time : 2023-01-17 ~ 2023-01-20 (4 half days) ,
On site Location : HG 115
URL/ZOOM location: https://unibe-ch.zoom.us/j/69257057684?pwd=dnNJaDRWcmtyQjRnbGVsTHc3UTkxUT09
On site Location : HG 115
URL/ZOOM location: https://unibe-ch.zoom.us/j/69257057684?pwd=dnNJaDRWcmtyQjRnbGVsTHc3UTkxUT09
Training language: English
Material: https://github.com/guiwitz/MLCV
Participants : Max 25
Registraion : Mandatory
Coaches : Dr. Guillaume Witz
Prerequisites : Laptop
Certificate : Certificate for full training attendance, 2 ECTS for successful project
Participants : Max 25
Registraion : Mandatory
Coaches : Dr. Guillaume Witz
Prerequisites : Laptop
Certificate : Certificate for full training attendance, 2 ECTS for successful project
About Lecturer
Guillaume is an imaging specialist working for the Science IT Support and the Microscopy Imaging Center. He has a background from EPFL and Harvard.
https://www.scits.unibe.ch/about_us/people_metadata/dr_witz_guillaume
https://www.scits.unibe.ch/about_us/people_metadata/dr_witz_guillaume