Hands-on NeuroAI

DozentIn: Dr. Sushrut Thorat , Prof. Dr. rer. nat. Tim Christian Kietzmann

Veranstaltungstyp: Blockseminar

Ort: 50/E08: Dienstag, 18.11.2025 10:00 - 11:00, Freitag, 21.11.2025, Freitag, 05.12.2025 13:00 - 17:00, Samstag, 06.12.2025 09:00 - 17:00, 50/E07: Samstag, 22.11.2025 09:00 - 17:00

Zeiten: Termine am Dienstag, 18.11.2025 10:00 - 11:00, Freitag, 21.11.2025 13:00 - 17:00, Samstag, 22.11.2025 09:00 - 17:00, Freitag, 05.12.2025 13:00 - 17:00, Samstag, 06.12.2025 09:00 - 17:00, Ort: 50/E08, 50/E07

Beschreibung: Course Description:
NeuroAI research requires us to both engineer new ML systems motivated by cognitive science / neuroscience theory and interpret their operation towards furthering and testing those theories. This is a hard endeavor as it requires expertise in many fields - cognitive/systems (neuroscience)science and machine learning. This block course aims to give students a flavor of the NeuroAI research cycle by introducing them to two problems recently tackled in the field: the role of recurrence in object recognition and the role of modulatory feedback in continual learning. After introduction, you will reproduce a substantial chunk of the methods and results from two central papers in those domains. By the end, you will have acquired invaluable intuitions in systems-design and interpretability required to carry out future research in NeuroAI.
Current situation re: coding, you will have to rely on your personal computers for training and deploying small neural networks. By small I mean smaller than ResNet-18 (try training a ResNet-18 on MNIST on your computer - try PyTorch - if that works you have more than enough computational power for the course).

Learning Objectives:
By the end of this course, students will:
Objective 1 - Understand the fundamental assumptions and intuitions of NeuroAI research
Objective 2 - Understand how to build toy ML systems to begin testing computational theories
Objective 3 - Understand how to run interpretability analysis on these trained systems

Prerequisites:
You have coded up and trained a simple neural network in PyTorch or Tensorflow. You have basic understanding of linear algebra. It would be helpful, but not necessary, if you have completed the ML4CCN course.

Course Format:
· Delivery Method: Fully in-person
· Type of Contact: Email the primary lecturer, Sushrut Thorat (sthorat@uos.de) in case of emergency/personal queries, else post your questions to Stud-IP.
· Selection Process: None

Assessment and Grading:
Each of the 4 coding sessions will contribute to the final grade. You will write a final report that summarizes further, previously unreported, analysis on one of the two systems, which will also contribute towards the final grade.

Important Dates:
· Exam signup deadline: 24 March, 2026
· There is no final exam - the completion of the coding sessions and the final report will decide the grade.

Required Texts and Materials or further Resources: None

Will this class be offered again/regularly?: No

zur Veranstaltung in Stud.IP