LIMITATIONS ON COMPLEXITY OF RANDOM LEARNING NETWORKS.
Randomly connected networks can be made adaptive, and thus able to "learn." Signal-to-noise considerations are shown to limit the maximum initial complexity which can learn. A higher order of complexity may be possible in multilayered structures which learn layer-by-layer; or if learning is possible during construction. Perception-like devices would appear not to be operative if of a high order of complexity.
Computers↗