BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:From Data to Models to Understanding: Evaluating Neural Latents an
 d Finding Decision Boundaries in Recurrent Neural Networks (RNNs) - Kabir 
 Dabholkar\, Technion - Israel Institute of Technology
DTSTART:20251029T160000Z
DTEND:20251029T170000Z
UID:TALK237811@talks.cam.ac.uk
CONTACT:124819
DESCRIPTION:Advances in machine learning have unlocked access to increasin
 gly rich computational models of cognition and its underlying neural dynam
 ics. This richness brings with it several challenges of different kinds. I
  will discuss two specific challenges and ways to address them. Model eval
 uation: ensuring that models fit to neural data align with the true underl
 ying dynamics\, to which we do not have direct access. I will show how the
  model’s few-shot generalisation – its ability to predict held-out par
 ts of the data from a few examples – helps quantify this match. This app
 roach selects models that capture the full richness of the data without 
 ‘inventing’ extraneous features. Model analysis: many dynamical models
  exhibit multistability\, marked by decision boundaries (separatrices) in 
 their state space\, which are hard to locate—especially in high dimensio
 ns. We introduce a Koopman‐theory‐driven neural network that learns a 
 scalar function vanishing on the separatrix\, and demonstrate its use on s
 imple systems and RNNs to design optimal perturbations for crossing these 
 boundaries and to make predictions for the outcome of optogenetic stimulat
 ions.
LOCATION:CBL Seminar Room\, Engineering Department\, 4th floor Baker build
 ing
END:VEVENT
END:VCALENDAR