BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Lazy-rich learning - Dr Alexander Rivkind\; Ishan Kalburge
DTSTART:20251118T131500Z
DTEND:20251118T124500Z
UID:TALK240919@talks.cam.ac.uk
CONTACT:124819
DESCRIPTION:Lazy-rich learning regime dichotomy is a crucial principle und
 erlying learning theory in both biological and artificial agents. Neural n
 etworks in the lazy regime are characterized by minimal weight changes\, f
 ast learning\, and high-dimensional representations at convergence corresp
 onding to kernel regression with the Neural Tangent Kernel\, whereas rich 
 representations are characterized by lower-dimensional feature learning wi
 th slow learning and larger (feature) gradients. We will first briefly rev
 iew these concepts as presented in Farrell et al. (https://www.sciencedire
 ct.com/science/article/pii/S0959438823001058) and then discuss evidence of
  rich neural representations in humans and macaques trained to perform con
 text-dependent decision-making.\nWe will then review a recent ICML'25 pape
 r by Chou et al. (https://arxiv.org/pdf/2503.18114) that demonstrates theo
 retically and empirically that the laziness-richness of  learning regime c
 an be evaluated using experimentally accessible metrics of representationa
 l geometry rather than via probing individual neurons\, synapses or featur
 es.
LOCATION:CBL Seminar Room\, Engineering Department\, 4th floor Baker build
 ing
END:VEVENT
END:VCALENDAR