A Computational Study of Reconstruction (from Partial Data) and Anticipation Capabilities of an Associative Neural Net with Large Stored Data-Base

I simulated a large Hopfield neural net which had the signum instead of sigmoid activation function so that it could be naturally physically implemented, e.g. in spin systems. It has been used in computational simulations in order to analyze the following capabilities of processing very large and complex data sets (e.g., protein-structure data-bases): 1. completion of patterns; 2. recognition of patterns; 3. prediction of unknown parameters; 4. a.nticipation. While for tasks 1-3 we use a memory-ba,se of previouslylearned examples using "a,ssociations", ta.sk 4 is equivalent to case 3 re-interpreted for temporal (or timeseries) prediction, i.e. prediction of unknown future parameter-values (instead of unknown present ones). For tasks 3 and 4 it is concluded that a generalization of the model used in simulation, like phase-Hebb processing or quantum-like information dynamics, if more promising. Data-structure conditions for success of tasks l-4 are discussed in a complex "real-life" example.