MSM's versus NN's

More about MSM's versus NN's

The picture shows (rather vaguely) two columns of op amps. Each output of the first column is connected via a weighting resistor to the summing junction of each of the op amps in the second column. This forms the analog version of a neural network. The individual gains (weighting resistors) are trainable in order to construct the model (e.g. a LLM -- large language model). In the digital version, the weighting resistors are replaced by matrix multiplications. i.e. (inputs) <-- (outputs) +.* (weightings). Thus training can easily adjust the weighting matrix. This is normally done on a GPU card, whic is designed for parallel matrix operations (SIMD or MMX). In a MSM (Markov State Machine), as used in ::SHE+ILA::, the op amps are replaced by virtual ALU's, implemented as states or coprocesses (threads), and the weightings are replaced by probabilities of a given branch between states being taken. (Cf quantum physics). This is sort of like a serial version of a NN (neural network) since the stimuli (e.g. prompt strings) propagate through the network of states step-by-step rather than in parallel (cf. RS232 versus Centronics Parallel). MSM technology has however been successfully used in spell-checkers, word prediction, pattern/speech recognisers, compilers (e.g. ::SHE+ILA::), and some language models. E&OE.