BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Connectome-based Echo State Networks - James McAllister (Ulster)
DTSTART:20250516T130000Z
DTEND:20250516T134500Z
UID:TALK232441@talks.cam.ac.uk
CONTACT:Daniel Kornai
DESCRIPTION:The Echo State Network (ESN) framework is an efficient recurre
 nt neural network paradigm of importance for both neuroscience and machine
  learning. It is quick and efficient to train\, and also has been suggeste
 d as a possible model of brain function. It is not fully known how network
  structure influences ESN functionality\, dynamics\, and robustness. We us
 ed biological networks to study this\, compared with randomly initialised 
 ESNs.\nIn both biological and artificial neural network contexts\, neurons
  and clusters demonstrate functional specificity. We asked if synapse-reso
 lution connectome-based ESNs demonstrate functional specificity/generality
  at the neural level. We used the larval Drosophila melanogaster connectom
 e\, which exhibits a hierarchical modular structure according to type\, cl
 ass\, and function. We built connectome-based ESNs from this and compared 
 neural specificity metrics between connectome and equivalent random networ
 ks across tasks in memory\, decision-making\, and time-series prediction. 
 Connectome ESNs contain smaller subsets of task-selective neurons\, while 
 random networks exhibit more distributed\, “general” groups. We tested
  the interpretation of these metrics by systematically pruning nodes based
  on their measure of engagement\, finding that connectome ESNs maintain pe
 rformance more robustly across pruning. Finally\, we investigated structur
 al features of the networks\, uncovering correlations between task-relevan
 ce and characteristics such as recurrence\, node degree\, and biological c
 ell-type annotations. We find that correlations are consistent across conn
 ectome and conventional ESNs\, but that connectome ESNs exhibit stronger c
 orrelations.\nThese findings indicate that biologically-inspired connectiv
 ity can enable sparsely selective and compact neural networks\, which may 
 reduce energy consumption and optimise robustness. The approaches and metr
 ics used in this research may suggest a way to initialise better performin
 g\, more efficient and robust ESNs\, as well as provide insight into indiv
 idual node feature importance in biological networks.
LOCATION:CBL Seminar Room\, Engineering Department\, 4th floor Baker build
 ing
END:VEVENT
END:VCALENDAR