Modern Hopfield Networks (MHNs) are a form of dense associative memory that also underpins modern attention mechanisms. According to a new arXiv preprint, MHNs that use linear kernels run into trouble when memory load is high: their attractor basins overlap and they produce spurious states, which limits their capacity.
The paper, titled "Phase Kernel Lifts Capacity of Dense Associative Memory," proposes a phase-based kernel to address this limitation. The abstract suggests that this kernel 'lifts' capacity, presumably by separating attractor basins and reducing spurious states, though the full details are not available in the abstract.
Since the source is a single preprint and the abstract is truncated, this digest can only report the problem and the proposed direction. No independent verification or comparison with other approaches is available.