BPCPPDYSOE21 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 25: Nonlinear Dynamics 2 - organized by Azam Gholami (Göttingen)

DY 25.2: Vortrag

Dienstag, 23. März 2021, 11:20–11:40, DYc

Damage-Resilient Computation in Spiking Neural Networks — •Fabio Schittler Neves, Georg Börner, and Marc Timme — Chair for Network Dynamics, Institute for Theoretical Physics & Center for Advancing Electronics Dresden (cfaed), TU Dresden, Dresden, Germany

Networks of spiking neurons with inhibitory coupling exhibit reconfigurable k-winner-take-all computations via changes to a single parameter [1], robustly determining the k strongest out of N analog inputs. Such partial rank ordering of signals provides a natural basis for computing arbitrary functions. Moreover, computations are completed within a few spikes (~k), thus requiring low power. Here we show that such networks are strongly resilient with respect to failure or removal of neural units. We develop strategies for immediate function recovery that work even after damage to an extremely large number of units. These networks exhibit two forms of resilience: first, the loss of less than N-k units do not translate in any change in dynamics, as the N-k neurons receiving the weaker inputs never spike, thus never contribute to any collective network dynamics; second, the systems provide great flexibility through symmetric coupling, because any unit in the network can functionally replace any other. Suitably interacting inhibitory neural networks may provide resilient and flexible analogue computations at low power and offer attractive solutions where unit repair or replacements are economically or practically infeasible, for example in autonomous and remote computing.

[1] F. S. Neves & M. Timme, IEEE Access 8:179648 (2020).