Architecture

[elias]

If I understand correctly, the CM-1K chip has 64 clusters of 16
neurons, totaling 1024 parallel coupled neurons. But is neuron to be
understood in conventional manner ie. as a function of the form
y=f(w_0p_0+w_1p_1+...+w_ip_i)? If so, what is the transfer function,
f, and can the neurons be coupled in some sort of multible layer
architecture?

[Robin D. Knight]

Hi Elias,

The CogniMem chip does not implement a traditional NN.  All of the NNs at a certain context (127 maximum) use either a Radial Basis Function (RBF) or a K-Nearest Neighbor (KNN) algorithm to categorize an input pattern to the nearest neuron's influence field (i.e. category).

The input pattern is up to 256 bytes. As you train neurons, the input pattern plus its category are presented to the chip. The chip's decision space gets segmented into influence fields, defined by several min/max parameters, RBF or KNN covering algorithm, and the distance metric (L1 or Lsup). Where the 256 byte pattern falls could positively feed back on a neuron if it is of the same category, or shrink its influence field if there is a category mismatch.

All neurons at a given context operate in parallel. To gain NN layers, the output of one context along with user define pattern info can be fed to the next context.

I hope this help,

-Robin Knight

--
Robin Knight
RoadNarrows LLC
www.roadnarrows.com
ph: +1 800.275.9568
fax: +1 800.275.9534

elias wrote:

If I understand correctly, the CM-1K chip has 64 clusters of 16
neurons, totaling 1024 parallel coupled neurons. But is neuron to be
understood in conventional manner ie. as a function of the form
y=f(w_0p_0+w_1p_1+...+w_ip_i)? If so, what is the transfer function,
f, and can the neurons be coupled in some sort of multible layer
architecture?


__________ Information from ESET NOD32 Antivirus, version of virus signature database 3336 (20080807) __________

The message was checked by ESET NOD32 Antivirus.

http://www.eset.com