Seminar aus Maschinellem Lernen und Data Mining
This seminar (TUCaN module number: 20-00-0102-se) covers advanced topics in machine learning. If you want to participate in the seminar, you have to register in TUCaN before Tuesday, 20.04.2020 23:55.
We will use Moodle to organize paper/review allocation and submission of files. Please sign up for this seminar here: https://moodle.informatik.tu-darmstadt.de/course/view.php?id=930
We will update this web page regularly.
Learning Objectives
In this seminar, students improve their ability to 1. understand the problem/scientific question addressed by a publication, 2. identify key contributions, 3. gather additional content from other publications that is helpful to understand/present the publication, 4. present the addressed problem/scientific question, the idea/solution, performed experiments, and the gained scientific knowledge to an audience and answer related questions, and 5. summarize and critically assess the quality of scientific papers.
Content
In this seminar, we plan to focus on two different machine learning topics: Deep Continuous-Discrete Machine Learning and Explainable AI.
Research in Deep Continuous-Discrete Machine Learning (also known as neural/symbolic machine learning) aims at combining recently successful continuous/neural deep learning with discrete/symbolic machine learning approaches. Exemplary research topics are integration of symbolic knowledge into neural networks, extraction of knowledge from neural networks, development of hybrid architectures, neural program synthesis, and neural computers.
Research in Explainable AI aims at developing methods that help humans to analyze, inspect, visualize, interpret, and understand complex AI systems such as deep neural networks. Exemplary research topics are learning what kind of explanations are helpful for humans, visualizing neural network behavior, trust in AI systems, and self-explanatory models / model-based interpretability.
In case you want to recap basics of deep learning, here are some resources:
- LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).
- Schmidhuber, J. Deep Learning in neural networks: An overview. Neural Networks 61, 85–117 (2015).
- Goodfellow, I., Bengio, Y. & Courville, A. Deep Learning. (2016).
Prerequisites
In this seminar, we discuss recent and advanced topics in machine learning and deep learning. Hence, having prior knowledge of machine learning and deep learning such as basic knowledge of rule learning, decision trees, neural networks, attention mechanism, and explainability is highly recommended.
Participation is limited to a fixed number of students (The exact number is to be determined). In case we have more students than papers/slots, we have to use random selection to select the students that can participate in the seminar. Students who did not sign up for the seminar in TUCaN before Tuesday, 20.04.2020 23:55 cannot be considered.
Organization
In this seminar, each student 1. prepares a talk for a topic/paper, 2. presents the topic/paper to the participants of the seminar, 3. answers questions during/after the talk, 4. writes a scientific review for a paper, and 5. actively participates in the question/answer sessions. Each task will be relevant for grading.
The language used in the seminar will be English. The topics for the talks and the scientific reviews will be assigned in the kick-off meeting.
Kick-off Meeting / Regular Meetings
All meetings will take place on Tuesdays at 17:10-18:50. Since physical meetings are not possible, we will use a video conference software for all meetings. More information can be found in Moodle.
The kick-off meeting will take place on Tuesday, 21.04.2020 at 17:10-18:50.
The first regular meeting will take place on Tuesday, 12.05.2020. The last regular meeting will take place on Tuesday, 14.07.2020.
Talks
The students are expected to give a talk divided into two parts. The first part (not more than10 minutes) should explain the problem/research question addressed by the paper and explain required background knowledge to understand the problem/research question. Explaining the background knowledge can require to use other resources such as other publications, books, blog posts, youtube videos etc. The second part (not more 20 minutes) should explain the solution/idea proposed by the paper and how claims/hypothesis made in the paper are validated (i.e. the experiments). The talk is followed by feedback, questions, and discussions.
Although each topic is typically associated with a single paper, the point of the talk is not to exactly reproduce the entire contents of the paper, but to communicate the key ideas of the methods that are introduced in the paper. Thus, the content of the talk should exceed the scope of the paper, and demonstrate that a thorough understanding of the material was achieved.
All publications should be freely available on the internet. Note that some paper sources such as Springer link often only works on campus networks (sometimes not even via VPN). If you cannot find a paper, contact us.