[nengo-user] About the model "RBM Deep Belief Network for Visual Digit Recognition"

[Terry Stewart]

Hello Zonghua Gu,

The approach we use for converting algorithms into spiking neurons is
called the "Neural Engineering Framework".  The core idea was
originally laid out in "Neural Engineering: Computation,
Representation, and Dynamics in Neurobiological Systems" by Chris
Eliasmith and Charlie Anderson (2003).  There's a quick summary
article here http://compneuro.uwaterloo.ca/files/publications/stewart.2012d.pdf

This approach allows you to make spiking neurons approximate a given
algorithm, as long as that algorithm is expressed in terms of vectors
and functions on those vectors.  Since an RBM (and, indeed, any
connectionist model) can be thought of this way, it is straightforward
to convert an RBM to spiking neurons simply by replacing each "neuron"
in the RBM with a small population of spiking neurons.

Given this methodology, we have developed software to automate this
conversion process.  This is called "Nengo", and it's what is being
used by the example you linked to.  The "digit.py" file you were
looking at is how we specify the algorithm we want the spiking neurons
to approximate.  If you run that script in Nengo, you will get a
network consisting entirely of spiking leaky-integrate-and-fire
neurons that approximates the given RBM.  In this particular case,
we're using 3 spiking neurons to replace each node in the first 3
layers of the RBM, and 10 spiking neurons in the last layer (these are
the N and N2 parameters in that script).

Does that help?
Terry





On Wed, Mar 25, 2015 at 2:43 AM, Zonghua Gu <zonghua at gmail.com> wrote:
> We built a FPGA for simulation of Spiking NN, and are looking for
> models as test applications. We hope to use the digit recognition
> example here http://models.nengo.ca/node/28. But from reading the
> paper, and the provided digit.py file:
>
> Tang Y, Eliasmith C. Deep networks for robust visual
> recognition[C]//Proceedings of the 27th International Conference on
> Machine Learning (ICML-10). 2010: 1055-1062.
>
> It seems it is not a spiking neural network. But the documentation
> says "A spiking neuron model for digit recognition, created by
> training an RBM Deep Belief Network on the MNIST database, then
> converting the resulting model to spiking neurons via Nengo." Could
> you please explain how the conversion is done?
>
>
> --
> Zonghua Gu
> Zhejiang University
> _______________________________________________
> nengo-user mailing list
> nengo-user at ctnsrv.uwaterloo.ca
> http://ctnsrv.uwaterloo.ca/mailman/listinfo/nengo-user