A Wasserstein Geometric Framework for Hebbian Plasticity

The real contribution here is not just a new formalism but a separation architecture: the framework proposes that what a network 'knows' (latent memory state as a probability measure) and what is physically observable (synaptic weights) are distinct objects connected by geometric projection. That split has practical consequences for how you would ever inspect or intervene on such a system.

The closest thread in recent coverage is the fixed-point analysis in 'Stability and Generalization in Looped Transformers' from April 16, which also asks how internal dynamics stabilize into observable outputs, just from a different angle. Both papers are working on the same underlying problem: characterizing what a network's internal state actually is and when it behaves predictably. The nonlinear separation principle paper from the same day is also adjacent, deriving structural conditions for stability in recurrent networks. This Wasserstein paper sits upstream of both, offering a geometric vocabulary that could, in principle, inform how those stability conditions are stated, though no direct connection is claimed.

Watch whether any group applies the Tan-HWG projection explicitly to a concrete recurrent architecture and reports whether the latent-to-weight separation produces measurably different learning trajectories than standard Hebbian updates on a benchmark like Penn Treebank or a controlled associative memory task.