Reproduce “Exponential Dynamic Energy Network for High Capacity Sequence Memory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arjun Karuvally, Pichsinee Lertsaroj, Terrence J. Sejnowski, Hava T. Siegelmann
EDEN models sequence recall by combining a fast memory network with a slowly changing asymmetric delay signal that shifts which memory becomes stable next. Its analysis predicts exponential fixed-point capacity in feature-neuron count, versus linear scaling for a reference network; escape-time predictions match simulations with mean absolute error 5.96 feature-neuron timescale units. These results assume random binary memories, not real-world episodes.
Key insights
- Global sequential dynamics can emerge from local convergence on a slowly evolving energy landscape, unifying static energy principles with temporal behavior.
- A critical ratio of self-memory to cross-memory interaction strengths ("alpha_s" / "alpha_c") dictates a phase transition between static and dynamic memory regimes, offering control over sequence transitions.
- The exponential memory capacity stems from the softmax-activated hidden layer, which effectively separates memory contributions within the energy function.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.