Date
Mon, 27 May 2024
Time
14:00 - 15:00
Location
Lecture Room 3
Speaker
Edoardo Ponti
Organisation
University of Edinburgh
Recent advancements in machine learning have caused a shift from traditional sparse modelling, which focuses on static feature selection in neural representations, to a paradigm based on selecting input or task-dependent pathways within neural networks. 
In fact, the ability to selectively (de)activate portions of neural computation graphs provides several advantages, including conditional computation, efficient parameter scaling, and compositional generalisation. 
 
In this talk, I will explore how sparse subnetworks can be identified dynamically and how parametric routing functions allow for recombining and locally adapting them in Large Language Models.


 

Last updated on 9 May 2024, 9:38am. Please contact us with feedback and comments about this page.