On Unstable Fixed Points in Modern Continuous Hopfield Networks

Abstract: The recently introduced continuous Hopfield network (see Ramsauer et al.) exhibits large memorization capabilities, which manifest as attractive fixed points of its update rule -- a differentiable function consisting of two linear mappings composed with the scaled softmax function. The authors of the aforementioned work provide proofs for the existence and approximate position of such attractive fixed points. For the softmax function alone, the fixed point structure has been fully characterized in earlier work by P. Tiňo, from which it turns out that for sufficiently large scaling factors there are exponentially more unstable fixed points than attractive ones. In this work, we complement the findings of Ramsauer et al. by showing that, under natural geometric conditions on the vectors defining the continuous Hopfield network, unstable fixed points must occur, analogous to the findings of Tiňo. Our results show that, under these geometric conditions, continuous Hopfield networks necessarily admit additional unstable fixed points associated with higher-dimensional faces of the pattern polytope.