Vision
- I got an EPSRC grant several years to build one of these agents,
and when I got it, I gulped because I realised I had to build
a vision system in neurons. I didn't know much about vision.
- It turns out that it's not that hard to build simple vision systems.
The vision guys say simple vision is easy, it's complex vision
that's hard.
- What comes down the retina is a bunch of on-off and off-on
centre-surround sensors.
- I coded that up with neurons quite easily.
- That goes (via the LGN which I skipped) to V1. This is a bunch
of edge, angle and line detectors.
- I coded those up too.
- These then go to higher areas. I just put in some stuff
to recognise a few objects (pyramids, stalactites and doors).
- We also did some texture recognition.
- The newest version has colour in too.
- We've never even tried to publish this stuff in vision journals.
- It does give us somewhat biologically realistic visual input.
- I'm kind of interested in doing written word recognition, but that's
just an idea.
