Hierarchical Uncertainty Exploration via Feedforward Posterior Trees

Authors: Elias Nehme, Rotem Mulayoff, Tomer Michaeli

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

Reproducibility Variable Result LLM Response
Research Type Experimental We showcase our approach across diverse datasets and image restoration challenges, highlighting its prowess in uncertainty quantification and visualization. Our findings reveal that our method performs comparably to a baseline that hierarchically clusters samples from a diffusion-based posterior sampler, yet achieves this with orders of magnitude greater speed.
Researcher Affiliation Academia Elias Nehme Electrical and Computer Engineering Technion Israel Institute of Technology seliasne@campus.technion.ac.il Rotem Mulayoff CISPA Helmholtz Center for Information Security rotem.mulayof@gmail.com Tomer Michaeli Electrical and Computer Engineering Technion Israel Institute of Technology tomer.m@ee.technion.ac.il
Pseudocode Yes Algorithm 1 Tree Training
Open Source Code Yes Code and examples are available at our webpage. Our code is made publicly available on Git Hub.
Open Datasets Yes Handwritten Digits, Edges Shoes, and Human Faces. Figure 2 demonstrates posterior trees on inpainting the top 70% of handwritten digits from the MNIST dataset. ... We also applied posterior trees to the edges-to-shoes dataset taken from pix2pix [56, 57]. ... Next, we tested posterior trees on face images from the Celeb A-HQ dataset, using the split from Celeb A [58]. ... Here, we applied our method to a dataset of migrating cells, imaged in a spinning-disk microscope, simultaneously with two different fluorescent dyes (one staining the nuclei and one staining actin filaments) [59].
Dataset Splits No For both components, the step size was reduced by a factor of 10 if the validation loss stagnated for more than 10 epochs, and the minimum step size was set to 5e-6.
Hardware Specification Yes Runtime is reported as both the speed of a forward pass (sec) and the memory usage (GB) required to infer a single test image with a batch of 1 on an A6000 GPU.
Software Dependencies No We used the Adam optimizer [55] with β1 = 0.9, β2 = 0.999 for all experiments. For the U-Net predicting the output leaves, we used an initial step size of 0.001. ... One can obtain an exact solution q to Eq. (A5) using standard methods for convex optimization, e.g. algorithms in CVXPY [63].
Experiment Setup Yes We used the Adam optimizer [55] with β1 = 0.9, β2 = 0.999 for all experiments. For the U-Net predicting the output leaves, we used an initial step size of 0.001. For the MLP predicting leaf probability, however, we found it beneficial to use a smaller initial step size of 0.0002. ... For both components, the step size was reduced by a factor of 10 if the validation loss stagnated for more than 10 epochs, and the minimum step size was set to 5e-6. We used a batch size of 32 for 70 epochs for all tasks.