Abstract: 

This tutorial targets researchers, practitioners, and advanced students in machine learning who seek to reduce the computational and energy costs of training large neural networks without sacrificing performance. Participants will learn the theory, algorithms, and system-level aspects of dynamic sparse training (DST) across supervised learning, reinforcement learning, generative Artificial Intelligence (AI), and truly sparse implementations. The tutorial combines algorithmic foundations with practical demonstrations using open-source code and commodity hardware, offering an innovative end-to-end perspective that connects cutting-edge research to executable systems. By the end, attendees will understand DST’s performance-efficiency trade-offs, gain hands-on experience, and join a dedicated Slack community to continue discussions, share results, and collaborate on advancing DST toward sustainable, Green AI.

Organizers:

• Dr. Elena Mocanu, Assistant Professor, University of Twente, The Netherlands

• Jafar Badour, University of Twente, The Netherlands

• Dr. Qiao Xiao, Postdoctoral Researcher, Eindhoven University of Technology, The Netherlands

• Boqian Wu, University of Twente, The Netherlands; University of Luxembourg, Luxembourg

• Dr. Decebal Constantin Mocanu, Associate Professor, University of Luxembourg, Luxembourg

Slides:

• Sparse Training for (Un)Supervised Learning [link]

• Sparse Training for Reinforcement Learning [link]

• Sparse Training for Generative AI [link]

• Truly Sparse Pipelines for Sparse Training [link] 

Slack channel: https://join.slack.com/t/sparsetrainin-ayg7306/shared_invite/zt-3ns6mfkqq-TwUNKLw3IGNliRAOoECWEw 

Lab materials: 

• https://github.com/boqian333/Truly-Sparse-Training 

• https://github.com/boqian333/Truly-Sparse-Training/blob/master/assignment/AAAI%202026-assignment.md