 |
Artificial Intelligence |
Overview
- OLAT
- Exercises in https://ki.inf.th-bingen.de/jupyter.
- Final project in groups of 1-3.
- This is the first iteration of this course. Expect some friction.
Exercises
- All exercises will be in the form of Jupyter notebooks and posted in OLAT
References
- "Deep Learning", Ian Goodfellow, Yoshua Bengio, Aaron Courville
- www.deeplearning.ai
History
| 1950s | first mentioning of "machine learning"; perceptron; neural networks |
| 1960s | nearest neighbor; backpropagation; discovery of fundamental limits for neural networks |
| 1970s | AI winter |
| 1980s | rediscovery of backpropagation; reinforcement learning |
| 1990s | AI winter; support vector machines; recurrent neural networks |
| 2000s | Netflix prize; ImageNet |
| 2010s | Deep Learning boom |
Machine translation
Google Translate
Object recognition
Protein folding
https://commons.wikimedia.org/wiki/File:C12orf29_AlphaFold.png
Games
But: deep learning is no magic pixie dust
https://commons.wikimedia.org/wiki/File:Fig-X_All_ML_as_a_subfield_of_AI.jpg
Why now?
What makes a successful DL teams?
1. Agressive data acquisition
2. Use every opportunity for automation
3. Excel at data warehousing
Module Goals
- Know relevant DL models and where to apply them
- Implement and train DL models
- Find out why they are not working
- Able to fix it
In the exercises we will implement a lot from scratch to see how the details work
Project
- Teams of 1 - 3 people
- Any topic you like
- Deep learning based
- Sufficient data available
- "unsolved"
- Grade based on presentation and report
- Approach
- Challenges
- Learnings
Project schedule
- Form teams till 14.10.2022
- 10. & 11.11.2022: Present topics (8 min per team)
- 10.1.2023: Report deadline
- From 12.1.2023: Final presentations (25 min per team)
Please talk to me about your proposed topic before 4.11.
-
Think about:
- What makes your topic relevant / interesting?
- What data will you use? How will you get it?
- How do you plan to approach your problem?
Report format
- Make a website (e.g. with GitHub or GitLab Pages)
describing
- the problem
- the data and how you acquired it
- general approach, used models, training
- how well it worked
- interesting findings and learnings
- Use images, graphs, videos, ...
- Optional: make it interactive
A good report would make a good Blog post.
Grading criteria
| Understandability | 20% |
| Reproducability | 10% |
| Innovativeness | 10% |
| Analysis of results | 20% |
| Useage of DL methods | 20% |
| Quality of approach | 20% |
Some example data sources
Jupyter has TensorFlow and PyTorch installed and two GPUs for training.
If something is missing on the machine - let me know.
Be considerate with resource usage. Do not abuse!
You can also use your own machine or any other compute resources you have access to.
For example Google Colab
Local Python / Jupyter installation:
Anaconda