Next: Neural networks and characteristics
Up: AI Lecture 2
Previous: Top-down vs. bottom-up approaches
- Simplified model of neurons and synapses
- Parallel implementation
- Information and processing are distributed across the network
(hence graceful degradation) ... but problems -
- NNs contain only local control. In the brain, local activity is
constantly broadcast over short and long distances.
- Restrictions on architecture - eg. layers and hidden units
- Scaling problem
- Catastrophic interference of new with old learning
- Finally, the ``neurons'' are usually macroscopically defined
entities - could be hypotheses, goals, properties - see the Ralph
Morelli example. Also they are idealised or ``mathematically pure''.
Thus the neural model is still more of an inspiration than a modelling
at the level of mechanism.