The Current CABot3s
- Here's the gross topology of one CABot3. The game
here is a virtual environment.
- The other boxes are component systems, that largely consist
of simulated or emulated neurons with static synapses.
- In my old Java simulator the NLP component made use of short-term
plasticity for binding.
- The cognitive map component takes advantage of long-term plasticity
to store a really simple cognitive map. Verb learning used
reinforcement learning (via synaptic plasticity) to learn some
simple action preferences.
- Aside from vision, the subsystems in the more complete versions
collectively have tens of thousands of neurons.
- There are a lot more neurons in the vision system.
- There are typically a lot of time-to-simulate issues, with the system
slowing down, but becoming more effective as more neurons are
used. In SpiNNaker, the parallel issues are time-to-load.
- So, I have made the vision flexible to work through this trade off.
- There are around 20 equally sized subnets (e.g. input, 3x3 on-centre
off-surround receptor, and
a horizontal edge detector). If the size is 20x20, that's 8000 neurons.
If it's 100x100, that's 200,0000 neurons.
- Each of the subsystems is relatively simple.
- What I'd like is a way to have lots of people effectively developing
complex subsystems and agents.