# The NEF Algorithm¶

While Nengo provides a flexible, general-purpose approach to neural modelling, it is sometimes useful to get a complete look at exactly what is going on “under the hood”. The theory behind the Neural Engineering Framework is developed at length in Eliasmith & Anderson, 2003: “Neural Engineering”, and a short summary is available in Stewart, 2012: “A Technical Overview of the Neural Engineering Framework.

However, for some people, the best description of an algorithm is the code itself. With that in mind, the following is a complete implementation of the NEF for the special case of two one-dimensional populations with a single connection between them. You can adjust the function being computed, the input to the system, and various neural parameters. In it written in Python, and requires Numpy (for the matrix inversion) and Matplotlib (to produce graphs of the output):

# A Minimal Example of the Neural Engineering Framework
#
# The NEF is a method for building large-scale neural models using realistic
#  neurons.  It is a neural compiler: you specify the high-level computations
#  the model needs to compute, and the properties of the neurons themselves,
#  and the NEF determines the neural connections needed to perform those
#  operations.
#
# The standard software for building NEF models is Nengo (http://nengo.ca).
#  Nengo is a cross-platform Java application that provides both a drag-and-drop
#  graphical user environment and a Python scripting interface for
#  creating these neural models.  It has been used to model a wide variety of
#  behaviour, including motor control, visual attention, serial recall, action
#  selection, working memory, attractor networks, inductive reasoning, path
#  integration, and planning with problem solving.
#
# However, given the complexity of Nengo, and due to the fact that this is a
#  fairly non-standard approach to neural modelling, we feel it is also useful
#  to have a simple example that shows exactly how the NEF works, from
#  beginning to end.  That is the goal of this script.
#
# This script shows how to build a simple feed-forward network of leaky
#  integrate-and-fire neurons where each population encodes a one-dimensional
#  value and the connection weights between the populations are optimized to
#  compute some arbitrary function.  This same approach is used in Nengo,
#  extended to multi-dimensional representation, multiple populations of
#  neurons, and recurrent connections.
#
# To change the input to the system, change 'input'
# To change the function computed by the weights, change 'function'
#
# The size of the populations and their neural properties can also be adjusted
#  by changing the parameters below.
#
# This script requires Python (http://www.python.org/) and Numpy
#  (http://numpy.scipy.org/) to run, and Matplotlib (http://matplotlib.org/) to
#  produce the output graphs.
#
# For more information on the Neural Engineering Framework and the Nengo

import random
import math

#################################################
# Parameters
#################################################

dt = 0.001       # simulation time step
t_rc = 0.02      # membrane RC time constant
t_ref = 0.002    # refractory period
t_pstc = 0.1     # post-synaptic time constant
N_A = 50         # number of neurons in first population
N_B = 40         # number of neurons in second population
N_samples = 100  # number of sample points to use when finding decoders
rate_A = 25, 75  # range of maximum firing rates for population A
rate_B = 50, 100 # range of maximum firing rates for population B

# the input to the system over time
def input(t):
return math.sin(t)

# the function to compute between A and B
def function(x):
return x * x

#################################################
# Step 1: Initialization
#################################################

# create random encoders for the two populations
encoder_A = [random.choice([-1, 1]) for i in range(N_A)]
encoder_B = [random.choice([-1, 1]) for i in range(N_B)]

def generate_gain_and_bias(count, intercept_low, intercept_high, rate_low, rate_high):
gain = []
bias = []
for i in range(count):
# desired intercept (x value for which the neuron starts firing
intercept = random.uniform(intercept_low, intercept_high)
# desired maximum rate (firing rate when x is maximum)
rate = random.uniform(rate_low, rate_high)

# this algorithm is specific to LIF neurons, but should
# generate gain and bias values to produce the desired
# intercept and rate
z = 1.0 / (1 - math.exp( (t_ref - (1.0 / rate) ) / t_rc) )
g = (1 - z) / (intercept - 1.0)
b = 1 - g*intercept
gain.append(g)
bias.append(b)
return gain, bias

# random gain and bias for the two populations
gain_A, bias_A = generate_gain_and_bias(N_A, -1, 1, rate_A[0], rate_A[1])
gain_B, bias_B = generate_gain_and_bias(N_B, -1, 1, rate_B[0], rate_B[1])

# a simple leaky integrate-and-fire model, scaled so that v=0 is resting
# voltage and v=1 is the firing threshold
def run_neurons(input, v, ref):
spikes = []
for i in range(len(v)):
dV = dt * (input[i] - v[i]) / t_rc  # the LIF voltage change equation
v[i] += dV
if v[i] < 0:
v[i] = 0                        # don't allow voltage to go below 0

if ref[i] > 0:                      # if we are in our refractory period
v[i] = 0                        #   keep voltage at zero and
ref[i] -= dt                    #   decrease the refractory period

if v[i] > 1:                        # if we have hit threshold
spikes.append(True)             #   spike
v[i] = 0                        #   reset the voltage
ref[i] = t_ref                  #   and set the refractory period
else:
spikes.append(False)
return spikes

# measure the spike rate of a whole population for a given represented value x
def compute_response(x, encoder, gain, bias, time_limit=0.5):
N = len(encoder)   # number of neurons
v = [0] * N        # voltage
ref = [0] * N      # refractory period

# compute input corresponding to x
input = []
for i in range(N):
input.append(x * encoder[i] * gain[i] + bias[i])
v[i] = random.uniform(0, 1)  # randomize the initial voltage level

count = [0] * N    # spike count for each neuron

# feed the input into the population for a given amount of time
t = 0
while t < time_limit:
spikes = run_neurons(input, v, ref)
for i, s in enumerate(spikes):
if s:
count[i] += 1
t += dt
return [c / time_limit for c in count]  # return the spike rate (in Hz)

# compute the tuning curves for a population
def compute_tuning_curves(encoder, gain, bias):
# generate a set of x values to sample at
x_values = [i * 2.0 / N_samples - 1.0 for i in range(N_samples)]

# build up a matrix of neural responses to each input (i.e. tuning curves)
A = []
for x in x_values:
response = compute_response(x, encoder, gain, bias)
A.append(response)
return x_values, A

# compute decoders
import numpy
def compute_decoder(encoder, gain, bias, function=lambda x:x):
# get the tuning curves
x_values, A = compute_tuning_curves(encoder, gain, bias)

# get the desired decoded value for each sample point
value = numpy.array([[function(x)] for x in x_values])

# find the optimal linear decoder
A = numpy.array(A).T
Gamma = numpy.dot(A, A.T)
Upsilon = numpy.dot(A, value)
Ginv = numpy.linalg.pinv(Gamma)
decoder = numpy.dot(Ginv, Upsilon) / dt
return decoder

# find the decoders for A and B
decoder_A = compute_decoder(encoder_A, gain_A, bias_A, function=function)
decoder_B = compute_decoder(encoder_B, gain_B, bias_B)

# compute the weight matrix
weights=numpy.dot(decoder_A, [encoder_B])

#################################################
# Step 2: Running the simulation
#################################################

v_A = [0.0] * N_A       # voltage for population A
ref_A = [0.0] * N_A     # refractory period for population A
input_A = [0.0] * N_A   # input for population A

v_B = [0.0] * N_B       # voltage for population B
ref_B = [0.0] * N_B     # refractory period for population B
input_B = [0.0] * N_B   # input for population B

# scaling factor for the post-synaptic filter
pstc_scale = 1.0 - math.exp(-dt / t_pstc)

# for storing simulation data to plot afterward
inputs = []
times = []
outputs = []
ideal = []

output = 0.0            # the decoded output value from population B
t = 0
while t < 10.0:
# call the input function to determine the input value
x = input(t)

# convert the input value into an input for each neuron
for i in range(N_A):
input_A[i] = x * encoder_A[i] * gain_A[i] + bias_A[i]

# run population A and determine which neurons spike
spikes_A = run_neurons(input_A, v_A, ref_A)

# decay all of the inputs (implementing the post-synaptic filter)
for j in range(N_B):
input_B[j] *= (1.0 - pstc_scale)
# for each neuron that spikes, increase the input current
# of all the neurons it is connected to by the synaptic
# connection weight
for i, s in enumerate(spikes_A):
if s:
for j in range(N_B):
input_B[j] += weights[i][j] * pstc_scale

# compute the total input into each neuron in population B
# (taking into account gain and bias)
total_B = [0] * N_B
for j in range(N_B):
total_B[j] = gain_B[j] * input_B[j] + bias_B[j]

# run population B and determine which neurons spike
spikes_B = run_neurons(total_B, v_B, ref_B)

# for each neuron in B that spikes, update our decoded value
# (also applying the same post-synaptic filter)
output *= (1.0 - pstc_scale)
for j, s in enumerate(spikes_B):
if s:
output += decoder_B[j][0] * pstc_scale

print t, output
times.append(t)
inputs.append(x)
outputs.append(output)
ideal.append(function(x))
t += dt

#################################################
# Step 3: Plot the results
#################################################

x, A = compute_tuning_curves(encoder_A, gain_A, bias_A)
x, B = compute_tuning_curves(encoder_B, gain_B, bias_B)

import pylab
pylab.figure()
pylab.plot(x, A)
pylab.title('Tuning curves for population A')

pylab.figure()
pylab.plot(x, B)
pylab.title('Tuning curves for population B')

pylab.figure()
pylab.plot(times, inputs, label='input')
pylab.plot(times, ideal, label='ideal')
pylab.plot(times, outputs, label='output')
pylab.title('Simulation results')
pylab.legend()
pylab.show()