A Learning Neural Cognitive Architecture
- What we'd really like is a neural learning Cognitive Architecture.
- Soar has chunking. ACT-R changes association strengths.
- We've used learning before. It's Hebbian, but seems quite
incomplete.
- Learning is always on. Soar often turns it off because it has
too many chunks.
- Standard architectures rarely run for more than a day.
- We'd like neural systems that run for months (or years) learning
all the time.
- We use Hebbian learning.
- We use reinforcement learning.
- This is all in the closed loop, but there may be other mechanisms.
- We haven't used neural and synaptic growth and death.
- Are there neurotransmitters for shutting down (turning on learning) in
areas?
