| Biologically Inspired Temporal
Evolutionary Neural Circuits |
| |
| Reza Derakhshani |
| |
| Biological neural networks have always motivated creation of new
artificial neural network models. One of the more complicated problems
in temporal neural network has been the incorporation of short and
long-term memory and most importantly design of efficient training
algorithms. Short-term memory (STM) usually takes the form of multiple
delayed copies of signals and thus provides a limited temporal history
of events similar to FIR filters (parallel forward delays in the diagram),
whereas the synaptic connection strengths as well as delayed feedback
loops (IIR form, see the delayed feedback loops from the output to
the input node in the figure) constitute longer-term memories (LTM).
I am developing a general evolutionary temporal neural network framework
(GETnet). This is a step towards general intelligent systems that
can be applied to a broad range of problems. GETnet incorporates nonlinear
moving average/ autoregressive nodes that are trained both through
modified gradient search and evolutionary programming, mimicking a
mixture of Lamarckian and Neo-Darwinian evolutionary mechanisms. The
ability to evolve arbitrary adaptive time-delay connections gives
GETnet the ability to find novel answers to many classification and
system identification tasks that include multidimensional time varying
inputs. |
| |
 |
| |
| Figure: Example GETnet structure which includes short-term memory
through parallel forward delays as well as delayed feedback loops
from the output to the input node which constitute longer-term memories. |
