Previous 
Up 
Next
Within this project we have developed an ANS simulator ("ANSim")
which is "all-neural" i.e. with adapting units, print-now signals,
gating of connections etc. in addition to the standard ANN
operations. It makes it possible to build both perceptual and
associative functions including learning and selforganization together
with motor pattern generation and control, in the same ANS and to let
these different systems interact. The basic building blocks for
specifying multi-network structures are populations, i.e. sets of
units with common properties, activation function etc. (possibly with
some distribution in some properties) and projections, i.e. bundles of
connections between a prepopulation and a postpopulation or
postprojection.
A population has spiking units with a passive membrane time constant
and with adaptation (which is important to have e.g. in CPG:s). There
are linear, exponential, and logarithmic transfer functions. The
connections can be delayed and they can integrate over time, i.e. be
fast or slow etc. Projections may be n:m or 1:1 configured and fixed
or modifiable. Sparse connectivity is not yet included but should be
handled efficiently for scaling to networks with large numbers of
units. There is an interface to the environment in the form of sensors
and actuators. It is set up to be run as a controller process, in
parallel with e.g. a robot simulator, under SIMON
, a
simple multi-processing tool developed within SANS.
This simulator exists in a beta version and a lot more remains to be done with it. It is not yet general enough to allow implementation of any neural paradigm or mixing of paradigms. At present, its core is Bayesian neural networks aimed at unsupervised and reinforcement learning and it will also implement self-organization which fits nicely in this context. Other important activities concern improving the user interaction as well as parallel and distributed implementation.
Next