Reservoir

Description

A reservoir is a pool of units with random connections. The connections are such that the activity of the network is “at the edge of chaos”. This concept forms the basis of a particular type of neural networks called liquid-state-machines (LSM).

• liquid state machine
• input
• read-out
• LSM with LIF neurons?

Implementation

• preliminary operations:

  import blabla
  import pyNCS
  
  nsetup = pyNCS.NeuroSetup('my_setuptype.xml', 'my_setup.xml')
  

• create a population:

  number_of_units = 500
  res = pyNCS.Population('', '')
  res.populate_by_number(nsetup,
                         'my_chip',
                         'my_neuron',
                         number_of_units)
  

• connect the units:

  C_res = pyNCS.Connection(res, res, 'excitatory0', fashion='random_all2all')
  

• create input:

  inp = pyNCS.Population('', '')
  inp.populate_by_id(nsetup,
                     'my_sequencer',
                     'my_neuron',
                     range(5, 10))
  

• connect the input to the reservoir:

  C_inp = pyNCS.Connection(inp, res, 'excitatory1') # default fashion is one2one
  

• make the input spike:

  pattern1 = inp.soma.spiketrains_poisson(random(len(inp)), duration=500)
  

• prepare the hardware:

  nsetup.chips[res.neuronblock.neurochip.id].loadBiases('biases/reservoir.biases')
  # the following is equivalent to the previous statement
  res.neuronblock.neurochip.loadBiases('biases/reservoir.biases')
  
  nsetup.mapping.write() # connections where automatically appended
  

• unleash hell:

  # stimulus lasts for 500ms but we want to record more, say 5s
  out = nsetup.stimulate(pattern1, tDuration=5000)
  

• plot (with monitors is a lot easier and faster!):

  # the output
  out[out.soma.channel].raster_plot()
  # external input and recurrent input
  imshow(out[out.synapses.channel].firing_rate(50)) # 50ms time-bin
  # input stimulus
  pattern1[inp.soma.channel].raster_plot()
  

Parameters

The parameters for a good reservoir are... ?