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..

Brain-like Variational Inference

Authors: Hadi Vafaii, Dekel Galor, Jacob L. Yates

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, i P-VAE outperforms both standard VAEs and Gaussian-based predictive coding models in sparsity, reconstruction, and biological plausibility, and scales to complex color image datasets such as Celeb A. i P-VAE also exhibits strong generalization to out-of-distribution inputs, exceeding hybrid iterative-amortized VAEs.
Researcher Affiliation Academia 1Redwood Center for Theoretical Neuroscience, UC Berkeley
Pseudocode Yes Algorithm 1 Iterative Generation Procedure for i P-VAE
Open Source Code Yes Code, data, and model checkpoints are available here: https://github.com/hadivafaii/Iterative VAE
Open Datasets Yes We train and evaluate models on two datasets: (1) whitened 16x16 natural image patches from the van Hateren dataset [100], used to assess convergence and reconstruction sparsity trade-offs; (2) MNIST [101], used for reconstruction, classification, and OOD generalization tests. ... In appendix C.9, we demonstrate practical utility by scaling i P-VAE to high-dimensional color images (Celeb A, CIFAR-10, tiny Image Net)
Dataset Splits Yes For our comparisons, we utilized the official implementations of sa-VAE [92], ia-VAE [91], and P-VAE [50]. Across models, we maintained consistent train/validation splits and hyperparameters unless otherwise specified.
Hardware Specification Yes training the models presented in Fig. 3 with Ttrain = 16 requires approximately three hours per model on an NVIDIA A6000 GPU. ... Our complete set of experiments, including all hyperparameter sweeps and model variants, was conducted over the course of approximately one week using six NVIDIA A6000 GPUs (48GB VRAM each).
Software Dependencies Yes We thank the developers of the software packages used in this project, including Py Torch [130], Num Py [131], Sci Py [132], scikit-learn [133], pandas [134], matplotlib [135], and seaborn [136].
Experiment Setup Yes For the reconstruction-sparsity analysis (Fig. 4), we swept across Ttrain {8, 16, 32} and β values proportional to each Ttrain, with factors {0.5, 0.75, 1.0, 1.25, 1.5, 2.0, 3.0, 4.0}. ... For the van Hateren dataset, we used a learning rate of 0.002, a batch size of 200, and trained for 300 epochs. MNIST [221] models used identical hyperparameters but were trained for 400 epochs.