H Dean,

NuPIC is not a general purpose neural circuit emulator.  It is a set of code
that implements a particular neural model.  If you like the principles of
neurobiology used  in this model it might be a good starting point for what
you want to do.  NuPIC is part of a long term goal to model the neocortex.
Not any particular neocortex, but general principles of neocortex such that
we can build many varieties of neocortex. 

 

At the heart of NuPIC is a model of a layer of neocortical cells, what we
call CLA, or cortical learning algorithm.  The neuroscience principles used
in this model are neurons with active dendrites, dendritic spikes that
depolarize the cell, the growth and decay of synapses as the primary mode of
learning, a micro-column architecture as a means of learning high order
sequences, sparse activations, and a few more.  The CLA primarily implements
a sequence memory that exhibits many desirable properties.  The way I view
it is the CLA is a fundamental building block of cortex a couple of steps up
from a single neuron.  It is the only model I know that incorporates active
dendritic properties, micro-columns, and typical connectivity in service of
high order sequence memory.  It is a good model of a layer of cells.  You
don't need to know all the neuroscience to use NuPIC but you will have to
learn the concepts.

 

Everything in the model was included because they are necessary for the
function we were trying to achieve, a distributed high order sequence memory
of sparse distributed representations.  There are many neuroscience details
we left out or modified for convenience, but all these decisions were
consciously made with knowledge of the neuroscience.  "Biological
plausibility" is not a term I like.  I think of the neuroscience as a set of
constraints.  We carefully choose which of those constraints are necessary
for machine intelligence and which are not.

 

If you want to model at a higher level, the caudate nucleus for example,
without using sparse distributed representations, sequence memory, etc. then
I don't think NuPIC is going to work for you.  However, I hope you give
NuPIC a careful look.  I believe the principles embodied in the CLA will be
used in all future brain models, I believe many of the principles apply
universally.

 

Jeff

 

From: nupic [mailto:[email protected]] On Behalf Of Dean Horak
Sent: Thursday, October 31, 2013 4:29 PM
To: [email protected]
Subject: [nupic-dev] Using NuPIC as an engine for WBE

 

I'm in the beginning stages of planning a project to perform whole brain
emulation and am considering using NuPIC as the engine. Unlike the BBP/BBP,
my plan is to focus on biological plausibility at the macroscale level of
nuclei, rather than concern for plausibility at the meso or neuronal level
for the sake of computation requirements. This is similar to the idea behind
Spaun, in that they are shooting for functional realism rather than
biological fidelity. 

I started building my models using the Nengo platform (which is used by
Spaun), but because it still shoots for relative biological plausibility at
the neuronal level, it's computational demands are simply too great for my
purpose (and my limited hardware). 

Therefore, my intent is to take my models (beginning with a simplified
caudate nucleus, simple sensory and motor cortices, rudimentary thalamus and
possible a basic hippocampus) and implement them on top of the NuPIC or
similar engine. I envision this as multiple distinct networks, each
interconnected, as in the thalami-cortical loop. Once I'm able to get this
simplified model performing correctly, I plan on building out additional
structures and refining the existing ones, as well as their connectivity in
an iterative fashion guided by experiment.

My question to the NuPIC group is, do you envision any architectural
limitations in NuPIC that might be of concern in using it in this manner, or
any thoughts on why NuPIC may or may not be a good fit for this project?

I've been toying with NuPIC since it's first release, but only as a
replacement for simple ANN applications, but I've just recently subscribed
to the maillist so if this is not the correct venue for a question such as
this, kindly let me know and I'll direct it elsewhere.

Thanks,
Dean Horak

_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to