This compiler will be used to program the ORCA neuromorphic processor using a high level Network Descript Language (NDL) such as brian2 or teili. Ultimately we would like to build a proper backend. Before doing this I would like to speak to James Knight and/or Thomas Nowotny as they developed the GeNN backend. I believe it ultimately breaks down to a code generation which converts the brian2's NDL into a c++/cuda code for GeNN and FPGA code for ORCA.

In the first iteration I will simply build a text generator function which extracts all necessary information from a TeiliNetwork or Network object, similar to teili2ctxctl as provided in teili tools. In contrast to the DYNAP interface ORCA needs more information about the TeiliNetwork as it has less limitations and higher flexibility and higher weight precision. A second important difference is that ORCA features on-chip learning so we don't need to transfer learned weights, but rather the same initialization parameters for the random distribution and the probability of connection, rather than exact indices.

Ultimately, we can also provide equations to ORCA, but for now the model is inherently fixed, so that the teili2orca 1.0 focuses on connectivity and general NDL rather than equations.

We need to grab the network object and extract the following information.

• Total number of neurons aka neuron count across population
• List/dict of neuron popluation
• Matching list/dict of population sizes
• List of populations (pre & post) which are connected via synapses
  • Sign of the synapse
  • Static vs. plastic
• Neuron & Synapse parameter dictionaries

In order to specify networks you either need to install brian2 or teili. Both network descriptions work out of the box and can be converted to an ORCA friendly format.

To install speed simply clone the repository and use pip to install the module by pointing it to the folder containing the setup.py.

pip install speed/

Alternatively, you can set your $PYTHONPATH pointing to where you cloned the repository to.

For more detailed and functional example please refer to speed/tutorials/convert_network_tutoral.py, You can find more tutorials at speed/tutorials.

import os

from speed.teili2orca import Speed
from teili import TeiliNetwork

Net = TeiliNetwork()
# Define your network

Net.run(10*second, report='text')


converted_model = Speed(Net)
converted_model.print_network()
converted_model.save_to_file(filename='orca_net.p',
                             directory=os.path.expanduser('~'))

When printing the network structure one expect an output similar to:

n_pop
   test_WTA__n_exc :  50
   test_WTA__n_inh :  12
s_pop
   test_WTA__s_exc_exc :  ['test_WTA__n_exc', 'test_WTA__n_exc']
   test_WTA__s_exc_inh :  ['test_WTA__n_exc', 'test_WTA__n_inh']
   noise_syn :  ['poissongroup', 'test_WTA__n_exc']
   test_WTA__s_inp_exc :  ['test_WTA__spike_gen', 'test_WTA__n_exc']
   test_WTA__s_inh_exc :  ['test_WTA__n_inh', 'test_WTA__n_exc']
n_params
   test_WTA__n_exc
     Vthr :  -50.4 mV
     EL :  -55. mV
     Iexp :  0. A
     Iconst :  0. A
     Cm :  281. pF
     Inoise :  0. A
     gL :  4.3 nS
     Iadapt :  0. A
     Vres :  -70.6 mV
     refP :  2. ms
   test_WTA__n_inh
     Vthr :  -50.4 mV
     EL :  -55. mV
     Iexp :  0. A
     Iconst :  0. A
     Cm :  281. pF
     Inoise :  0. A
     gL :  4.3 nS
     Iadapt :  0. A
     Vres :  -70.6 mV
     refP :  2. ms
s_total 2069
n_total 62
s_params
   test_WTA__s_exc_exc
     w_plast :  1
     baseweight :  1. nA
     kernel :  0. A s^-1
     tausyn :  5. ms
   test_WTA__s_exc_inh
     w_plast :  1
     baseweight :  1. nA
     kernel :  0. A s^-1
     tausyn :  5. ms
   noise_syn
     w_plast :  1
     baseweight :  1. nA
     kernel :  0. A s^-1
     tausyn :  5. ms
   test_WTA__s_inp_exc
     w_plast :  1
     baseweight :  1. nA
     kernel :  0. A s^-1
     tausyn :  5. ms
   test_WTA__s_inh_exc
     w_plast :  1
     baseweight :  1. nA
     kernel :  0. A s^-1
     tausyn :  5. ms
s_tags
   test_WTA__s_exc_exc
     plastic :  False
     p_connection :  0.376
     mean :  1.0
     std :  0.0
   test_WTA__s_exc_inh
     plastic :  False
     p_connection :  0.715
     mean :  1.0
     std :  0.0
   noise_syn
     plastic :  False
     p_connection :  0.0004
     mean :  1.0
     std :  0.0
   test_WTA__s_inp_exc
     plastic :  False
     p_connection :  0.02
     mean :  1.0
     std :  0.0
   test_WTA__s_inh_exc
     plastic :  False
     p_connection :  1.0
     mean :  1.0
     std :  0.0