To simplify creating connection matrices for high-dimensional ensembles, you can use three additional parameters in the nef.Network.connect() function: weight, index_pre, and index_post. weight specifies the overall gain on the connection across all dimensions, and defaults to 1. For example:
net.make('A',100,3) net.make('B',100,3) net.connect('A','B',weight=0.5) # makes a transform matrix of # [[0.5,0,0],[0,0.5,0],[0,0,0.5]]
Note that the system by default assumes the identity matrix for the connection.
If you don’t want the identity matrix, and would prefer some other connectivity, specify index_pre and index_post. These indicate which dimensions in the first ensemble should be mapped to which dimensions in the second ensemble. For example:
net.make('A',100,3) net.make('B',100,1) net.connect('A','B',index_pre=2) # makes a transform matrix of # [[0,0,1]] net.make('A',100,1) net.make('B',100,3) net.connect('A','B',index_post=0) # makes a transform matrix of # [,,] net.make('A',100,4) net.make('B',100,2) net.connect('A','B',index_pre=[1,2]) # makes a transform matrix of # [[0,1,0,0],[0,0,1,0]] # which makes B hold the 2nd and 3rd element of A net.make('A',100,4) net.make('B',100,3) net.connect('A','B',index_pre=[1,2],index_post=[0,1]) # makes a transform matrix of # [[0,1,0,0],[0,0,1,0],[0,0,0,0]] # which makes B hold the 2nd and 3rd element of A # in its first two elements
To make the inputs to neurons noisy, you can specify an amount of noise and a noise frequency (how often a new noise value is sampled from the uniform distribution between -noise and +noise). Each neuron will sample from this distribution at this rate, and add the resulting value to its input current. The frequency defaults to 1000Hz:
Here is how you can convert an input to provide a randomly changing value, rather than a constant:
net.make_fourier_input('input', dimensions=1, base=0.1, high=10, power=0.5, seed=0)
This will produce a randomly varying input. This input will consist of random sine waves varying from 0.1Hz to 10Hz, in 0.1Hz increments. The random number seed used is 0.
You can set an ensemble to be simulated as spiking neurons, rate neurons, or directly (no neurons). The default is spiking neurons:
One common usage of direct mode is to quickly test out algorithms without worrying about the neural implementation. This can be especially important when creating algorithms with large numbers of dimensions, since they would require large numbers of neurons to simulate. It can often be much faster to test the algorithm without neurons in direct mode before switching to a realistic neural model.
When using direct mode, you may want to decrease the number of neurons in the population to 1, as this makes it much faster to create the ensemble.
When building models that represent large numbers of dimensions, it is sometimes useful to break an ensemble down into sub-ensembles, each of which represent a subset of dimensions. Instead of building one large ensemble to represent 100 dimensions, we might have 10 ensembles that represent 10 dimensions each, or 100 ensembles representing 1 dimension each.
The main advantage of this is speed: It is much faster for the NEF methods to compute decoders for many small ensembles, rather than one big one.
However, there is one large disadvantage: you cannot compute nonlinear functions that use values in two different ensembles. One of the core claims of the NEF is that we can only approximate nonlinear functions of two (or more) variables if there are neurons that respond to both dimensions. However, it is still possible to compute any linear function.
We create an array by specifying its length and (optionally) the number of dimensions per ensemble (the default is 1):
You can also use all of the parameters available in nef.Network.make() to configure the properties of the neurons.
The neurons parameter specifies the number of neurons in each ensemble, not the total number of neurons!
The resulting array can be used just like a normal ensemble. The following example makes a single 10-dimensional ensemble and a network array of 5 two-dimensional ensembles and connects one to the other:
net.make_array('A',neurons=100,length=5,dimensions=2) net.make('B',neurons=500,dimensions=10) net.connect('A','B')
When computing nonlinear functions with an array, the function is applied to each ensemble separately. The following computes the products of five pairs of numbers, storing the results in a single 5-dimensional array:
net.make_array('A',neurons=100,length=5,dimensions=2) net.make('B',neurons=500,dimensions=5) def product(x): return x*x net.connect('A','B',func=product)
To simplify the manipulation of matrices, we have added a version of JNumeric to Nengo. This allows for a syntax similar to Matlab, but based on the NumPy python module.
To use this for matrix manipulation, you will first have to convert any matrix you have into an array object:
a=[[1,2,3],[4,5,6]] # old method a=array([[1,2,3],[4,5,6]]) # new method
You can also specify the storage format to be used as follows:
a=array([[1,2,3],[4,5,6]],typecode='f') # valid values for the typecode parameter: # 'i' int32 # 's' int16 # 'l' int64 # '1' int8 # 'f' float32 # 'd' float64 # 'F' complex64 # 'D' complex128
The first important thing you can do with this array is use full slice syntax. This is the [:] notation used to access part of an array. A slice is a set of three values, all of which are optional. [a:b:c] means to start at index a, go to index b (but not include index b), and have a step size of c between items. The default for a is 0, for b is the length of the array, and c is 1. For multiple dimensions, we put a comma between slices for each dimension. The following examples are all for a 2D array. Note that the order of the 2nd and 3rd parameters are reversed from matlab, and it is all indexed starting at 0:
a # the first row a[0,:] # the first row a[:,0] # the first column a[0:3] # the first three rows a[:,0:3] # the first three columns a[:,:3] # the first three columns (the leading zero is optional) a[:,2:] # all columns from the 2nd to the end (the end value is optional) a[:,:-1] # all columns except the last one (negative numbers index from the end) a[::2] # just the even-numbered rows (skip every other row) a[::3] # every third row a[::-1] # all rows in reverse order a[:,::2] # just the even-numbered columns (skip every other column) a[:,::-1] # all columns in reverse order a.T # transpose
With such an array, you can perform element-wise operations as follows:
c=a+b # same as .+ in matlab c=a*b # same as .* in matlab b=cos(a) # computes cosine of all values in a # other known functions: add, subtract, multiply, divide, remainder, power, # arccos, arccosh, arcsinh, arctan, arctanh, ceil, conjugate, imaginary, # cos, cosh, exp, floor, log, log10, real, sin, sinh, sqrt, tan, tanh, # maximum, minimum, equal, not_equal, less, less_equal, greater, # greater_equal, logical_and, logical_or, logical_xor, logical_not, # bitwise_and, bitwise_or, bitwise_xor
You can also create particular arrays:
arange(5) # same as array(range(5))==[0,1,2,3,4] arange(2,5) # same as array(range(2,5))==[2,3,4] eye(5) # 5x5 identity matrix ones((3,2)) # 3x2 matrix of all 1 ones((3,2),typecode='f') # 3x2 matrix of all 1.0 (floating point values) zeros((3,2)) # 3x2 matrix of all 0
The following functions help manipulate the shape of a matrix:
a.shape # get the current size of the matrix b=reshape(a,(3,4)) # convert to a 3x4 matrix (must already have 12 elements) b=resize(a,(3,4)) # convert to a 3x4 matrix (can start at any size) b=ravel(a) # convert to a 1-D vector b=diag([1,2,3]) # create a diagonal matrix with the given values
Some basic linear algebra operations are available:
c=dot(a,b) c=dot(a,a.T) c=innerproduct(a,a) c=convolve(a,b)
And a Fourier transform:
The following functions also exist:
# argmax, argsort, argmin, asarray, bitwise_not, choose, clip, compress, # concatenate, fromfunction, indices, nonzero, searchsorted, sort, take # where, tostring, fromstring, trace, repeat, diagonal # sum, cumsum, product, cumproduct, alltrue, sometrue
The vast majority of the time, you can use these objects the same way you would a normal list of values (i.e. for specifying transformation matrices). If you ever need to explicitly convert one back into a list, you can call .tolist():
These functions are all available at the Nengo console and in any script called using the run command. To access them in a separate script file, you need to call:
from numeric import *