Learning
- So far I've largely overlooked learning.
- Learning is essential to the argument and to AI in general
- One of the points of the relative failure of Cyc is that
its really hard to code all of the knowledge; some type of
machine learning is needed.
- Fortunately, Neural Nets are ideal for learning.
- In particular, we make use of Hebbian learning, the
only type of learning that has extensive biological
evidence. (Neural and synaptic death and growth are
also backed.)
- The initial system should learn from the environment
preferably in an unsupervised mode.
- It is likely that areas closer to the IO interface
are learned first while later areas develop more
slowly.
- It's really essential to figure out learning and how it
relates to the neural net.
