González, M., Sánchez, Á., Dominguez, D., & Rodríguez, F. B. “Fine-Tuning of Patterns Assignment to Subnetworks Increases the Capacity of an Attractor Network Ensemble”. In International Work-Conference on Artificial Neural Networks. 2021, June, pp. 236-247. Springer, Cham.
It is known that dividing an attractor network into a set of subnetworks whose connectivity is equivalent to the attractor network from which they come, and therefore with the same computational cost, increases the system’s recovery capacity. This opens the possibility of optimizing the assignment of pattern subsets to the ensemble modules. The patterns subsets assignment to the network modules can be considered as a combinatorial optimization problem, where varied strategies (i.e. random vs. heuristic assignments) can be tested. In this work, we present a possible heuristic strategy driven by an overlap minimization in the subsets for assigning the patterns input to the modules of the ensemble attractor neural network. In terms of system pattern storage capacity, the assignment driven by the overlap minimization in each subset/module proved to be better than no specific assignment, i.e. distribution of patterns to modules randomly.