Nonlinear neural networks: Principles, mechanisms, and architectures

TLDR: An historical discussion is provided of the intellectual trends that caused nineteenth century interdisciplinary studies of physics and psychobiology by leading scientists such as Helmholtz, Maxwell, and Mach to splinter into separate twentieth-century scientific movements.

About: This article is published in Neural Networks.The article was published on 1988-01-01 and is currently open access. It has received 1586 citations till now. The article focuses on the topics: Competitive learning & Adaptive resonance theory.

Figures

FIGURE 7. Alphabet learning: Code learning in response to the first presentation of the first 20 letters of tlhe alphabet is shown. Two differel1lt vigilance levels were used, (a) p = .5 and (b) p = .8. Each row represents the total code that is learned after the letter at the left-hl~nd column of the row is presented at F,. Each column represents the critical feature pattern that is learned through time by the F2 node listed at the top of the column. The critical feature patterns do not, in general, equal the pattern exemplars which change them through learning. Instead, each critical feature pattern acts likE! a prototype for the entire siet of these exemplars, as well as for unfamiliar exemplars which share invariant properties with familiar exemplars. The simulation illustrates the "fast learning" calse, in which the altered L TM traces reach a new equilibrium in response to each new stimulus. Slow learning is more

FIGURE 9. (a) Learning in sensory-motor pathways is gated by a difference vector (DV) process which matches target position command (TPC) with present position command (PPC) to prevent incorrect associations from forming between eye-head TPCs and hand-arm TPCs; (b) A GO signal gates execution of

FIGURE 8. Circuit diagram of the back propagation model: In addition to the processing levels F" F2' F3' there are also I,evels F., Fs, F6' and F7 to carry out the computations which control the learning process. The transport of learned weights from the F2 -+ F3 pathways to the F. -+ Fs pathways show that this algorithm cannot represent a learning process in the brain. All feedback within the levels F, -+ F2 -+ F3 is expressed through learning signals which slowly change bottom-up adaptive weights but do not quickly reorganize the network's fast information processing.

Frequently asked questions

Q1. What are the contributions in "Nonlinear neural networks: principles, mechanisms, and architectures" ?

The remainder of the article describes results about continuous-nonlinear models: