% IMPORTANT: The following is UTF-8 encoded.  This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.

@INPROCEEDINGS{Keup:878012,
      author       = {Keup, Christian and Tobias, Kühn and Dahmen, David and
                      Helias, Moritz},
      title        = {{T}ransient chaotic dimensionality expansion by recurrent
                      networks},
      reportid     = {FZJ-2020-02581},
      year         = {2020},
      abstract     = {Cortical neurons communicate with spikes, which are
                      discrete events in time. Functional network models often
                      employ rate units that are continuously coupled by analog
                      signals. Is there a benefit of discrete signaling? By a
                      unified mean-field theory for large random networks of rate
                      and binary units, we show that both models have identical
                      second order statistics. Their stimulus processing
                      properties, however, are radically different: We discover a
                      chaotic submanifold in binary networks that does not exist
                      in rate models. Its dimensionality increases with time after
                      stimulus onset and reaches a fixed point that depends on the
                      synaptic coupling strength. Low dimensional stimuli are
                      transiently expanded into higher-dimensional representations
                      that live within the manifold. We find that classification
                      performance peaks when stimulus dimensionality matches the
                      submanifold dimension; typically within a single neuronal
                      time constant. Classification shows a high resilience to
                      noise that exceeds rate models by orders of magnitude. Our
                      theory mechanistically explains all these observations.These
                      findings have several implications. 1) Optimal performance
                      is reached with weaker synapses in discrete state networks
                      compared to rate models; implying lower energetic costs for
                      synaptic transmission. 2) The classification mechanism is
                      robust to noise, compatible with fluctuations in biophysical
                      systems. 3) Optimal performance is reached when each neuron
                      in the network has been activated only once; this
                      demonstrates efficient event-based computation with short
                      latencies. 4) The presence of a chaotic sub-manifold has
                      implications for the variability of neuronal activity; the
                      theory predicts a transient increase of variability after
                      stimulus onset. Our theory thus provides a new link between
                      recurrent and chaotic dynamics of functional networks,
                      neuronal variability, and dimensionality of neuronal
                      responses.},
      month         = {Feb},
      date          = {2020-02-27},
      organization  = {COSYNE, DENVER (USA), 27 Feb 2020 - 3
                       Mar 2020},
      cin          = {INM-6 / INM-10 / IAS-6},
      cid          = {I:(DE-Juel1)INM-6-20090406 / I:(DE-Juel1)INM-10-20170113 /
                      I:(DE-Juel1)IAS-6-20130828},
      pnm          = {574 - Theory, modelling and simulation (POF3-574) / PhD no
                      Grant - Doktorand ohne besondere Förderung
                      (PHD-NO-GRANT-20170405) / neuroIC002 - Recurrence and
                      stochasticity for neuro-inspired computation
                      (EXS-SF-neuroIC002) / MSNN - Theory of multi-scale neuronal
                      networks (HGF-SMHB-2014-2018)},
      pid          = {G:(DE-HGF)POF3-574 / G:(DE-Juel1)PHD-NO-GRANT-20170405 /
                      G:(DE-82)EXS-SF-neuroIC002 / G:(DE-Juel1)HGF-SMHB-2014-2018},
      typ          = {PUB:(DE-HGF)1},
      url          = {https://juser.fz-juelich.de/record/878012},
}