Contribution to a conference proceedings

FZJ-2021-00223

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Unfolding recurrence by Green’s functions for optimized reservoir computing

Nestler, S. (Corresponding author)FZJ* ; Keup, C.FZJ* ; Dahmen, D.FZJ* ; Gilson, M.FZJ* ; Rauhut, H. ; Helias, M.FZJ*

2020

Please use a persistent id in citations: http://hdl.handle.net/2128/26881

Abstract: Cortical networks are strongly recurrent, and neurons have intrinsic temporaldynamics. This sets them apart from deep feed-forward networks. Despite thetremendous progress in the application of feed-forward networks and their the-oretical understanding, it remains unclear how the interplay of recurrence andnon-linearities in recurrent cortical networks contributes to their function. Thepurpose of this work is to present a solvable recurrent network model that links tofeed forward networks. By perturbative methods we transform the time-continuous,recurrent dynamics into an effective feed-forward structure of linear and non-lineartemporal kernels. The resulting analytical expressions allow us to build optimaltime-series classifiers from random reservoir networks. Firstly, this allows us tooptimize not only the readout vectors, but also the input projection, demonstratinga strong potential performance gain. Secondly, the analysis exposes how the secondorder stimulus statistics is a crucial element that interacts with the non-linearity ofthe dynamics and boosts performance.

---

Contributing Institute(s):

1. Computational and Systems Neuroscience (INM-6)
2. Theoretical Neuroscience (IAS-6)
3. Jara-Institut Brain structure-function relationships (INM-10)

Research Program(s):

1. 574 - Theory, modelling and simulation (POF3-574) (POF3-574)
2. HBP SGA3 - Human Brain Project Specific Grant Agreement 3 (945539) (945539)
3. neuroIC002 - Recurrence and stochasticity for neuro-inspired computation (EXS-SF-neuroIC002) (EXS-SF-neuroIC002)
4. Advanced Computing Architectures (aca_20190115) (aca_20190115)
5. RenormalizedFlows - Transparent Deep Learning with Renormalized Flows (BMBF-01IS19077A) (BMBF-01IS19077A)
6. SDS005 - Towards an integrated data science of complex natural systems (PF-JARA-SDS005) (PF-JARA-SDS005)

Appears in the scientific report 2020

---

Database coverage:

---