BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:BMI Seminar // Friedemann Zenke - Learning World Models via Intern
 al Prediction: From Local Plasticity to Recurrent Circuits
DTSTART:20260903T111500
DTEND:20260903T121500
DTSTAMP:20261006T055113Z
UID:5189dd38650d831e0017677028b389f74d3e3e01672e2ff15f180a72
CATEGORIES:Conferences - Seminars
DESCRIPTION:Friedemann Zenke\nTo guide behavior\, the brain must extract o
 bject identities and their dynamics from entangled sensory streams\, and u
 se this information to anticipate what will happen next. How it builds suc
 h internal models from experience remains an open question. I will briefly
  recap earlier work showing that a local\, predictive form of Hebbian plas
 ticity allows neural networks to learn invariant object representations wi
 thout supervision\, using learning rules consistent with known properties 
 of synaptic plasticity.\n\nThe main part of the talk will focus on how pre
 dictive principles scale from invariance learning to circuit-level computa
 tion to yield internal world models. I will first introduce a theoretical 
 framework\, latent distribution matching (LDM)\, that casts self-supervise
 d learning as balancing predictive alignment with entropy-driven uniformit
 y in latent space. LDM unifies a broad class of existing self-supervised o
 bjectives\, including contrastive\, non-contrastive\, and joint-embedding 
 predictive approaches\, and yields identifiability guarantees for nonlinea
 r predictive models. Building on these principles\, I will present recurre
 nt predictive learning (RPL)\, a self-supervised joint-embedding predictiv
 e architecture that learns untangled representations of objects and their 
 dynamics directly from sensory streams\, without reconstruction or labels\
 , and that develops successor-like and abstract sequence representations r
 eminiscent of those found in visual and prefrontal cortex. Finally\, I wil
 l show how these ideas extend to model-based reinforcement learning throug
 h a reconstruction-free variant of Dreamer\, in which a JEPA-style predict
 or operating on continuous\, deterministic representations matches the per
 formance of reconstruction-based world models.\n\nTogether\, these results
  point toward an emerging framework in which internal prediction is the dr
 iver of plasticity. However\, for the brain to tap into such powerful repr
 esentation learning\, local learning seems too limited and we have to reth
 ink plasticity mechanisms that act at the circuit level.\n 
LOCATION:SV 3615 https://plan.epfl.ch/?room==SV%203615
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR