...
Below is a summary of sound-alike word training used in Release 8 modifications. The general approach is apply weights as memory addresses, in order to locate "similar things" (words, image features, etc.) physically near each other in memory as opposed to numerically near each other in vector space. Localizing data makes sense in terms of biological evolution and efficiency; the human brain consumes 40 Watts and has nowhere near the energy, numerical accuracy, and EDAC to perform complex algorithms such as gradient descent.
...