Increasing Memory Capacity and Reducing Spurious States in Neural Networks by Introducing Coherent and Collective Firing

It is well known that higher-order Hopfield nets called multispin models can increase memory capacity to some extent by extending the direct product of spin states to more than second order. However, a group of neurons can then respond degenerately to different loaded patterns, resulting in many spurious states due to cross-talk effects. We present an idea to increase the number of attracting basins for patterns while suppressing the associated spurious states, by introducing coherent and collective firing in multispin groups. We numerically implement the method and test the number, stability, and basin size of the attractors thus created. Increasing the size of a group of coherent excitation suppresses spurious states, stabilizes loaded patterns, and dramatically increases the number of pattern attractors.