BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Multitasking and compositionality in brain and in neural networks 
 - Alexander Rivkind\; Ishan Kalburge
DTSTART:20250617T100000Z
DTEND:20250617T113000Z
UID:TALK233494@talks.cam.ac.uk
CONTACT:124819
DESCRIPTION:This talk will synthesize findings from two key papers that in
 vestigate the modular and compositional nature of neural computation. We w
 ill explore two distinct perspectives on how neural networks achieve flexi
 ble behavior in multitask settings by reusing learned computational primit
 ives.\n\nFirst\, we will review Ito et al. 2022 ("Compositional generaliza
 tion through abstract representations in human and artificial neural netwo
 rks":https://papers.neurips.cc/paper_files/paper/2022/file/d0241a0fb1fc9be
 477bdfde5e0da276a-Paper-Conference.pdf)\, who used fMRI and a highly compo
 sitional task to identify abstract representations (i.e. orthogonalization
 ) as a neural substrate for compositional generalization in humans. They d
 emonstrated that pretraining artificial neural networks (ANNs) on basic ta
 sk “primitives” induces similar abstract representations\, enabling ze
 ro-shot generalization and human-like performance.\n\nNext\, we’ll turn 
 to the study by Driscoll et al.\, 2024 ("Flexible multitask computation in
  recurrent networks utilizes shared dynamical motifs.":https://www.nature.
 com/articles/s41593-024-01668-6) In this work\, the authors show that prim
 itives of neural computation—like attractors and decision boundaries—c
 an arise naturally in a monolithic recurrent neural network\, without bein
 g explicitly engineered for. Authors support this by locating and tracking
  attractors in the network’s phase space\, clustering tasks based on the
 ir neural activity\, and demonstrating that simulated “lesions” select
 ively disrupt tasks in line with those clusters. We will take a critical l
 ook at these findings in our journal club\, discussing the methods’ appl
 icability as well as their strengths and limitations.
LOCATION:CBL Seminar Room\, Engineering Department\, 4th floor Baker build
 ing
END:VEVENT
END:VCALENDAR