Sharing Neurons
Why Not?
© 2007 Ben Margolis
What's this thing you
have about sharing neurons between nets?
-
I believe a shared-neuron
architecture to be more biologically feasible than generating large numbers of
randomized neurons every time we have something new to learn. see:
Observations on Human Learning and
Human Language
-
Sharing input neurons is a
convenience, and sharing output neurons enables multi-tiered MLPs with any
change to standard MLP architecture. see:
Sharing I/O Neurons, A Simple Implementation
Modular Multi-Tier MLP Design
-
With a minor modification to
standard MLP architecture we can create a particularly biologically feasible
multi-function, shared-neuron, overlapping-network design, which may be
able to apply knowledge from one task to a different task and actually learn
from the similarities of separately trained tasks, hopefully becoming
"smarter" in the process.
Neural Network with Simulated Dendritic Tree
The Ad-Hoc Neural Network
|