Paramagnetic unlearning in neural network models
Kazuo Nokura · Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics · 1996
We study unlearning in the paramagnetic phase of neural network models. After many unlearning steps at temperature T, changes of synaptic interactions are expressed by the paramagnetic correlation function. Taking the second order terms of \ensuremath{\beta}=${\mathit{T}}^{\mathrm{\ensuremath{-}}1}$, we derive the evolution equation and find that the Hopfield model evolves into the pseudo-inverse model in some parameter region. As a second initial condition, we study the Hopfield model, which has varying pattern weights. In this case, unlearning works on the large weight patterns and removes the monopoly of them. \textcopyright{} 1996 The American Physical Society.