Modeling Mismatch Negativity: Difference between revisions

From KIP Wiki
ā§¼kip-jumptonavigationā§½ā§¼kip-jumptosearchā§½
(Created page with "=== Overview === Mismatch negativity (MMN) refers to the modified evoked potential of recordings that occur upon an unexpected event. For example, in the auditory domain, a s...")
Ā 
No edit summary
Ā 
Line 7: Line 7:


== Developers ==
== Developers ==
Catherine Wancogne
Catherine Wacongne




Line 34: Line 34:
== Results of the ESS workshop, Oct 2011 ==
== Results of the ESS workshop, Oct 2011 ==


The model was presented at the ''Using the ESS + Neuromorphic hardware'' workshop in Dresden, Oct 4th, 2011 by Catherine Wancogne. The slides of the presentation can be found [https://brainscales.kip.uni-heidelberg.de/internal/wiki/images/5/53/Mismatch_Negativity.pdf here].
The model was presented at the ''Using the ESS + Neuromorphic hardware'' workshop in Dresden, Oct 4th, 2011 by Catherine Wacongne. The slides of the presentation can be found [https://brainscales.kip.uni-heidelberg.de/internal/wiki/images/5/53/Mismatch_Negativity.pdf here].





Latest revision as of 15:27, 19 April 2012

Overview

Mismatch negativity (MMN) refers to the modified evoked potential of recordings that occur upon an unexpected event. For example, in the auditory domain, a subject may "get used" to a repetivie sound, evoking a particular (e.g. EEG) response. If the sequence is interspersed by an unknown sound the evoked response shows a strong negative component as compared to the previous responses. Mismatch negativity can also occur for more complex sequences of sounds [Winkler 2004, Hughes 2001] and spans several modalities. This model focuses on the cause of MMN in the auditory cortex across several layers. The short-term memory that is necessary to provide a comparison is accounted for by a system of synfire chains. The stimulus enters in layer 4 and propagates to layers 2/3 and the synfire chains. The model also provides feedback across the two layers, modulating the input layer 4 during presentation of the next stimulus. A stimulus that has previously not been applied to the system is translated to an elevated signal in layers 2/3 and 4.

Developing institutions

INSERM-CEA

Developers

Catherine Wacongne


Model Requirements

  • AdEx, LIF neurons (originally: Izhikevich)
  • STDP synapses (originally: STDP with modulation)
  • background Poisson input (300 Hz, individual to each neuron)

Modifications for Execution on the ESS

  • STDP is currently not available on the ESS
    • split model execution into learning part, executed on software simulator (NeST, Neuron), and static part, executed on the ESS
    • weights are written to a file from the learning simulation, which is then used by the non-learning ESS simulation
  • Rate of background noise exceeds the bandwidth limits of the off-wafer network
    • use a set of low-rate Poisson spike sources and randomly connect several of them to each neuron, such that the total rate equals the required 300 Hz

Requests for ESS Extensions

  • Implement model of hardware STDP in the ESS
    • Learning simulation can be directly run on the ESS
  • Monitor pulse loss in off-wafer network


Results of the ESS workshop, Oct 2011

The model was presented at the Using the ESS + Neuromorphic hardware workshop in Dresden, Oct 4th, 2011 by Catherine Wacongne. The slides of the presentation can be found here.


  • An initial version of the model has been implemented in PyNN and thested on the ESS.
  • Massive Poisson Input needed, downscaling/compensation required
  • Reported problems with inhibition when running on ESS
    • BV: the problem arose from the fact, that there were neuron being excitatory and inhibitory at the same time. This is currently not supported by the mapping process but is no hardware limitation.