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..
Dense Associative Memory with Epanechnikov Energy
Authors: Benjamin Hoover, Zhaoyang Shi, Krishnakumar Balasubramanian, Dmitry Krotov, Parikshit Ram
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results show that LSR energy has significantly more local minima (memories) that have comparable log-likelihood to LSE-based models. |
| Researcher Affiliation | Collaboration | Benjamin Hoover IBM Research Georgia Tech Zhaoyang Shi Harvard Krishnakumar Balasubramanian UC Davis Dmitry Krotov IBM Research Parikshit Ram IBM Research |
| Pseudocode | Yes | Algorithm 1: Fixed Point Computation for the LSR Memory Retrieval |
| Open Source Code | Yes | The codebase is published in a Git Hub repository and contains necessary instructions to setup the environment and to recreate all results, down to the same seed used for training and sampling. |
| Open Datasets | Yes | To study this behavior, we use a VAE to encode MNIST and Tiny Imagenet [27] images into latent vectors that serve as the stored patterns for LSR and LSE energies (see fig. 4). |
| Dataset Splits | No | For MNIST, Ξ is obtained as follows: 24 random images from the MNIST training set are normalized to be [0, 1], rasterized into a 784-dim vector, and projected into a 10-dim latent space of a β-VAE trained according to the methods laid out below. These latents become our stored patterns ΞMNIST [0, 1]24 768 that we use to parameterize both the LSR energy and the LSE energy. The procedure for Tiny Imagenet [27] is similar. We randomly select 40 images from the dataset, each of shape (C,H,W)=(3,64,64). The paper does not provide explicit train/test/validation splits for its primary experiments evaluating the AM model. It randomly samples data points from existing datasets to serve as memories without specifying reproducibility of this sampling (e.g., random seed for selection of images) or traditional splits for evaluating the AM itself. |
| Hardware Specification | Yes | Each experiment was run on a system with access to 8x A40 GPUs each with 40GB of memory. |
| Software Dependencies | No | Experiments use both Py Torch [29] and JAX [30]. The paper mentions PyTorch and JAX but does not provide specific version numbers for these software libraries, which are necessary for reproducible descriptions. |
| Experiment Setup | Yes | Training proceeded for 50 epochs using a learning rate of 1e-3, the Adam optimizer, and a minibatch size of 128. The β of the β-VAE (distinct from the inverse temperature β used by the LSR energy) is set to 4. |