The Brain Does not have a Rigid Topology
- The brain does not have a rigid topology (e.g. Braitenberg)
- complementary unidirectional connections
- We've set up a topology that is rigid
- Given the way we normally set it up (distance biased connections),
can the weights be set?
- Is the way we normally set it up okay?
- Recursively up to 5 steps
- decreasing likelihood
- 1/4 are long-distance based (we've done lots of experiments
without this)
- Inhibitory connections are just random throughout the net
- Undoubtably, setting the connection weights after this would have
to be much more flexible than our current way.
- Might try .6 to neurons in your row, .4 to adjacent,
-.4 to adjacent but two, and -.2 to others. (All abiding by Gray's Law.)
- More interestingly, can this or a more flexible version be learned?
(I suspect it would be relatively easy.)
- In general, how can we convert these calculated nets to nets that
can be learned?
