Topic exploration part 1

Your task as a group is to study four topics in the group and produce a number of deliverables (see further down). At the end, each of you should be able to give a basic presentation about each of the topics to show that you meet the requirement of having basic knowledge in these areas.

Some resource pages

• Supervised Learning
• Unsupervised Learning
• Reinforcement Learning
• Deep Learning

Topics

The topics to be covered in the groups are defined below. Note that part of the job is to find the material for the topics. We provide some material (see above) but depending on what you know you might want to fill in some gaps after this or before.

• (Non-deep) ML basics
  • Why does one split data into three sets: Training, validation and test data.
  • What is validation by leave-one-out and K-fold cross validation?
  • Describe basic ideas and assumptions with the methods:
    • Linear regression
    • Logistic regression
    • K-nearest regression
    • Linear and Quadratic Discriminant Analysis (LDA, QDA)
    • Tree-based methods: classification trees, random forests, bagging
    • Support Vector Machines
    • Boosting
  • Demo one or more of these methods as they will be used in Music Taste Prediction. 
• (Deep) Neural Network basics
  • Neuron model
  • How to combine them into networks
  • Basic cost functions
  • How to learn the parameters using gradient descent.
  • Deep network challenges
  • Demo and explain the baseline provided for Deep Learning for Natural Language Processing
• Basic introduction to Reinforcement Learning
  • Describe problem type
    
  • Traditional vs deep RL
  • Give examples of where it works well and what is challenging
  • Show example(s)/demo(s)
• ML and Security, Privacy, FAT (Fairness, Accountability, Transparency), 6 topics to pick from, grouped by their relation to ML. You should pick one topic for the group.
  
  1. Machine learning for security (ML and Security)
     • What kinds of security problems can/cannot be helped by ML?
     • Provide an example for security by ML to explain what is done.
     • What are potential problems?
  2. Adversarial machine learning (ML and Security)
     • What kinds of attacks are there?
     • Give an example to show what is being attacked and how.
     • What countermeasures exist?
     • What is the difference to GAN (Generative Adversarial Networks)?
  3. Privacy-preserving machine learning (ML and Privacy)
     • What are the main privacy concerns when it comes to machine learning?* Give an example to illustrate the problem
     • What are potential consequences of privacy breaches?
     • Pick a proposed solution (for example, using cryptography and secure multi-party computation), explain how it works and its limitations.
  4. Differential Privacy (ML and Privacy)
     • What problem does it address?
     • How does it work?
     • Give an example to show how data is treated.
     • What are the costs and trade-offs when it comes to machine learning?
  5. Data anonymization/re-identification (ML and Privacy)
     • What does k-anonymity mean?
     • What problem does it address?
     • Give an example to show how data can be transformed to fulfill k-anonymity.
     • What concerns are addressed by l-diversity and t-closeness?
     • What are the effects on machine learning?
  6. FAT ML (fairness, accountability, transparency for machine learning)
     • What are the FAT concerns when it comes to machine learning?
     • Provide an example
     • What are some of the proposed countermeasures?

Deliverables

• For each of the topics
  • One 2h "teaching sessions" where the "student teacher" of a topic transfers the knowledge to the rest of the group. The sessions should be planned and announced so that the course teachers (Patric, Bo, Christian, Daniel or Juan Carlos) can listen in unannounced on some of them. These should be posted in this schedule 
    https://docs.google.com/spreadsheets/d/1qVvp1waZQwKTh_fsZPZPBS3I2EG3EHzsHWDl3IRVDQg/edit?usp=sharing Links to an external site.
  • We encourage the use of flipped classroom techniques, where some material is prepared before and read by "students" to prepare for the learning session, but we leave decision of the format to you.
  • We suggest that you use your group's Homepage in Canvas to collaborate on this assignment so that the material is seamlessly available to the course teachers as well. This page provides some info about the group.
• Presentation material for a 10-15 min lecture for each of the above topics that covers what you consider to be the important aspects in that topics. The presentation should include pointers to the sources of information that you used.
  • When possible the lecture should include a small demo of the main concepts, tools, etc in this topic. Feel free to reuse whatever you find on the material pages in this module and elsewhere.
  • The lecture should cover the basic material that you believe that everyone should know about the topic. One of the key things to discuss is why this topic is important to know about?
  • Connected to the presentation there should be a set of questions (to ask the audience) that capture the most important aspects of the topic. 
• One document that states
  • how the work was distributed with an explicit mentioning of who prepared what topic and is signed by everyone that the group believes has contributed fairly to the work. An evaluation of what you think of this form of learning.

Suggested way to work

Given that there are four topics and the groups consist of five people you can put more people on some topic. We leave it to you to do this planning. You would typically  prepare the lecture of the topics individually or in pairs and then present this to the other members of the group in 2h sessions. This would allow the "students" to give feedback on the lecture and material at the same time as they get to know the topic and the material so that in the end everyone has reached the goal of broad knowledge in these topics.

How will this be examined

• There will be an individual examination for the basic understanding of classical ML methods in Music Taste Prediction and on deep learning methods by completing one of both of Deep Learning for Natural Language Processing  and Food Recognition.
• Unannounced appearances by teachers at the teaching sessions.
• Uploaded presentation material will be assessed.

Answers to some questions asked

• Q: Why do we have 2h teaching sessions when we only have to prepare a 10-15min presentation for each topic?
  • A: The 2h session is where you (start to?) transfer your knowledge of the topic to the other students. This might contain more material than what makes it to the 10-15min final presentation . At the end of the assignment all in the group can deliver that 10-15min presentation and actually understand what they say, not just repeat the words of the teacher in the group or read the slide notes. This will take more than 10-15min.

Answers to some expected questions

• Q: We would like to pick some other topics  than the provided. Can we?
  • A: Sure, no problem as long as your group covers the four topics above as well and that everyone in the groups agree to learn about the additional topics that you add in addition to the mandatory ones. That is, we do not allow someone to go off on their own.
• Q: No one in my group wants to study topic X. Can we skip that?
  • A: No.
• Q: We have become fewer than five in our group. Can we skip some topic?
  • A: No, but we acknowledge that you do not have as much time at your hands and will take that into account in our assessment.