BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Jaeok Yi (KAIST)\, "Neural Networks from the Perspective of Physic
 s"
DTSTART:20260401T060000Z
DTEND:20260401T070000Z
DTSTAMP:20260507T054500Z
UID:indico-event-1254@indico.ibs.re.kr
DESCRIPTION:Despite the remarkable empirical success of deep learning\, a 
 comprehensive theoretical understanding of why and how neural networks lea
 rn remains a mystery. In this talk\, we discuss physics-inspired approache
 s to understanding neural networks. We present synaptic field theory\, a f
 ramework that reformulates the gradient descent dynamics of synaptic weigh
 ts as classical field dynamics in de Sitter spacetime\, constructing an ac
 tion whose metric naturally matches that of a universe with a positive cos
 mological constant. This framework faces a challenge related to the non-lo
 cality of the cost function. To address this issue\, we explore the idea o
 f promoting neurons to dynamical degrees of freedom. Leveraging properties
  of stochastic gradient descent\, the Lagrangian can be decomposed into a 
 data-independent bulk part and a data-dependent boundary part. This decomp
 osition is expected to separate the architectural structure from the stoch
 astic properties of neural networks\, enabling independent analysis of eac
 h. Through this line of research\, we aim to provide physicists with a fam
 iliar language to investigate the theoretical foundations of machine learn
 ing.\n\nhttps://indico.ibs.re.kr/event/1254/
LOCATION:CTPU Seminar room (Theory Bldg\, 4F)
URL:https://indico.ibs.re.kr/event/1254/
END:VEVENT
END:VCALENDAR