Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Exponential Dynamic Energy Network for High Capacity Sequence Memory
Authors: Arjun Karuvally, Pichsinee Lertsaroj, Terrence J. Sejnowski, Hava T. Siegelmann
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our work extends the energy paradigm to temporal memories by allowing the energy surface to change slowly with time. This approach was previously proposed experimentally in [34] and some computational properties studied in [35]. In contrast to the classical energy paradigm, the memories in the dynamic energy networks can lose or gain stability over time, resulting in stability in two timescales. In short timescales, the current memory is always stable, with the energy function guaranteeing convergence and robustness to noise. In longer timescales, the energy surface changes to create a new minimum, destroying the current minimum. The network state changes in response, resulting in stable transitions between memory states. Our analysis of the proposed dynamic energy network shows that (1) The network s dynamical behavior is well characterized by the short-timescale energy functions assembled piecemeal for long-timescale dynamical behavior, (2) The energy function provides a precise analytic computation for the time required to escape from a stable memory state and the conditions necessary to exhibit memory transitions, (3) The network capacity scales exponentially in the number of neurons, significantly outperforming existing sequence memory networks, (4) The network populations have biological implications, showing strong behavioral correlations to the activity of cells found in human episodic memory experiments. |
| Researcher Affiliation | Academia | Arjun Karuvally Salk Institute for Biological Studies EMAIL Pichsinee Lertsaroj University of Massachusetts Amherst Terrence J. Sejnowski Salk Institute for Biological Studies EMAIL Hava T. Siegelmann University of Massachusetts Amherst EMAIL |
| Pseudocode | No | The paper describes the mathematical model and its dynamics using equations and narrative, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The github repo for running the capacity experiments can be found at https://github.com/arjunkaruvally/EDEN_torch. |
| Open Datasets | Yes | EDEN is simulated to store and retrieve a simple sequence of 5 MNIST digits in numeric order. |
| Dataset Splits | No | To evaluate capacity, we ran simulations to estimate the probability of errors in retrieving single bits Pr h vi(te) ξ(µ) i 1 ϵ i for the fixed point error rate ϵ = 10 3. For each neuron setting N {10, 12,...32} and the number of memories from P {1, 2...2N}, the probability is estimated using Monte Carlo simulations. 100 seeds of memory initializations were taken with the memories sampled without replacement to avoid confusion in the retrieved memory sequence. After evaluating the single-bit error probability, the maximum number of memories to be stored is computed for an error rate of δ = 10 3. |
| Hardware Specification | No | The simulations were coded in Python and run in the Unity supercomputing cluster. |
| Software Dependencies | No | The paper mentions that simulations were coded in Python and uses a repository named 'EDEN_torch', implying PyTorch, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | For all numerical simulations of network state dynamics, we used the Euler integration procedure with a step size of 0.01. The memories in EDEN are defined as random binary vectors with each dimension of the memory in the model drawn from the Rademacher distribution Pr h ξ(µ) j = +1 i = Pr h ξ(µ) j = 1 i = 1/2. For Figure 2, 3, 5, the simulations were run with N = 100, αs = 0.98, αc = 1.0, Tf = 1.0, and Td = 20.0. To evaluate capacity, we ran simulations to estimate the probability of errors in retrieving single bits Pr h vi(te) ξ(µ) i 1 ϵ i for the fixed point error rate ϵ = 10 3. For each neuron setting N {10, 12,...32} and the number of memories from P {1, 2...2N}, the probability is estimated using Monte Carlo simulations. 100 seeds of memory initializations were taken. |