Inflationary Flows: Calibrated Bayesian Inference with Diffusion-Based Models

Authors: Daniela de Albuquerque, John Pearson

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate experimentally that inflationary flows indeed preserve local neighborhood structure, allowing for sampling-based uncertainty estimation, and that these models continue to provide high-quality generation under compression, even from latent spaces reduced to as little as 0.03% of the nominal data dimensionality.
Researcher Affiliation Academia Daniela de Albuquerque Department of Electrical & Computer Engineering School of Medicine Duke University Durham, NC 27708 daniela.de.albuquerque@duke.edu John Pearson Department of Neurobiology Department of Electrical & Computer Engineering Center for Cognitive Neuroscience Duke University Durham, NC 27708 john.pearson@duke.edu
Pseudocode Yes To integrate (83), we utilize either Euler s method for toy datasets and Heun s method (see Algorithm 1) for high-dimensional image datasets.
Open Source Code Yes Additionally, we provide entire code needed to reproduce results of paper in this repository [63].
Open Datasets Yes We performed two sets of experiments on two benchmark image datasets (CIFAR-10 [61] and AFHQv2 [62]...) ... CIFAR-10 [61]: MIT license AFHQv2 [62]: Creative Commons BY-NC-SA 4.0 license Toys [93]: BSD License. The NeurIPS checklist also explicitly states: "All datasets utilized are publicly available and we provide details on how to download and pre-process these data in our repository and in the appendices."
Dataset Splits Yes To assess how these two factors affect model performance, we performed two sets of experiments on two benchmark image datasets (CIFAR-10 [61] and AFHQv2 [62]...)... In Tables 1 and 2 we showcase Frechet Inception Distance (FID) scores [65] (mean 2σ over 3 independently generated sets of images, each with 50,000 samples) and round-trip integration mean squared errors (mean MSE 2σ over 3 randomly sampled sets of images, each with 10,000 samples) for each (d, IG) combination explored (Appendices B.5.1, B.5.2, B.5.4). From NeurIPS checklist: "Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes]"
Hardware Specification Yes training was performed in a distributed fashion using either 8 or 4 GPUs per each experiment (NVIDIA Ge Force GTX TITAN X, RTX 2080) in a compute cluster setting. Generation (FID) and round-trip (MSE) experiments were performed on single GPU (NVIDIA RTX 3090, 4090, A5000, A6000).
Software Dependencies No The paper mentions utilizing 'DDPM++ architecture, as implemented in [49]' and 'Toy datasets were obtained from scikit-learn [93]', but it does not provide explicit version numbers for general software dependencies like Python, PyTorch, or specific libraries used beyond these mentions.
Experiment Setup Yes For all cases, we used a learning rate of 10 5, batch size of 8192, and exponential moving average half-life of 50 104. ... Shown in Tables 8, 9 are our specific choices for the exponential inflation constant (ρ) and training duration (in 106 images Mimgs)... Table 7: CIFAR-10 & AFHQv2 Common Training Hyperparameters (Across All Schedules) Hyperparameter Name Hyperparameter Value Channel multiplier 128 Channels per resolution 2-2-2 Dataset x-flips No Augment Probability 12% Dropout Probability 13% Learning rate 10 4 LR Ramp-Up (Mimg) 10 EMA Half-Life (Mimg) 0.5 Batch-Size 512