On gauge freedom, conservativity and intrinsic dimensionality estimation in diffusion models

Authors: Christian Horvat, Jean-Pascal Pfister

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

Reproducibility Variable Result LLM Response
Research Type Experimental In section 5, we present a method for estimating the intrinsic dimensionality of the data manifold by analyzing the local variability when approaching the data manifold. Additionally, we provide empirical evidence that with a nonconservative vector field, the ID is not estimated correctly while using the derived method we can estimate the ID correctly if sθ is constrained to be conservative (not necessarily matching s exactly).
Researcher Affiliation Academia Christian Horvat Department of Physiology University of Bern christian.horvat@unibe.ch Jean-Pascal Pfister Department of Physiology Bern, Switzerland jeanpascal.pfister@unibe.ch
Pseudocode No No explicit pseudocode or algorithm blocks were found in the paper.
Open Source Code No The paper does not contain an explicit statement about the release of source code or a link to a code repository.
Open Datasets No The paper uses synthetic data (2-dimensional Gaussian, spheres, tori, swiss rolls) for its experiments, but does not provide access information for a publicly available or open dataset.
Dataset Splits No The paper does not specify dataset splits (e.g., train, validation, test percentages or counts) or reference predefined splits for reproducibility.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory, or cloud resources) used for running experiments are mentioned in the paper.
Software Dependencies No No specific software dependencies with version numbers (e.g., library names like PyTorch, TensorFlow, or specific solvers with versions) are mentioned in the paper.
Experiment Setup Yes We train a conservative and non-conservative diffusion model using f(x, t) = 0 and g(t) = 25t as drift and diffusion coefficient, respectively. The non-conservative diffusion model is simply an unconstrained neural network sθ(x, t) = ψθ where ψθ : R5 R>0 R5. The conservative version is sθ(x, t) = ||ψθ(x, t)||2 2 as suggested by Du et al. (2023). For the non-conservative diffusion model, we simply use a standard feed-forward neural network where we first embed the data into 100 dimensions and linearly transform it followed by a nonlinearity (first step). Further, we embed the resulting features into 200 dimensions, again linearly transform it followed by a non-linearity, and finally project back into the data dimensions (second step). We embed the time into 100 dimensions using a Gaussian-Fourier projection and add these embeddings to the features after the first step.