Differentiable Random Partition Models

Authors: Thomas Sutter, Alain Ryser, Joram Liebeskind, Julia Vogt

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach in three experiments, demonstrating the proposed DRPM s versatility and advantages. First, we apply the DRPM to a variational clustering task, highlighting how the reparametrizable sampling of partitions allows us to learn a novel kind of Variational Autoencoder (VAE, Kingma and Welling, 2014). By leveraging potential dependencies between samples in a dataset, DRPM-based clustering overcomes the simplified i.i.d. assumption of previous works, which used categorical priors (Jiang et al., 2016). In our second experiment, we demonstrate how to retrieve sets of shared and independent generative factors of paired images using the proposed DRPM. In contrast to previous works (Bouchacourt et al., 2018; Hosoya, 2018; Locatello et al., 2020), which rely on strong assumptions or heuristics, the DRPM enables end-to-end inference of generative factors. Finally, we perform multitask learning (MTL) by using the DRPM as a building block in a deterministic pipeline. We show how the DRPM learns to assign subsets of network neurons to specific tasks.
Researcher Affiliation Academia Thomas M. Sutter , Alain Ryser , Joram Liebeskind, Julia E. Vogt Department of Computer Science ETH Zurich Correspondence to {thomas.sutter,alain.ryser}@inf.ethz.ch
Pseudocode No The paper describes methods and procedures using prose and mathematical equations but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes We provide the code under https://github.com/thomassutter/drpm
Open Datasets Yes To assess the clustering performance, we train our model on two different datasets, namely MNIST (Le Cun et al., 1998) and Fashion-MNIST (FMNIST, Xiao et al., 2017)
Dataset Splits No The paper frequently mentions using 'test sets' for evaluation (e.g., 'on test sets of MNIST and FMNIST') and total training epochs, but it does not provide specific training/validation/test dataset splits (e.g., percentages, absolute counts, or references to predefined splits) for reproducibility.
Hardware Specification Yes All our experiments were run on RTX2080Ti GPUs.
Software Dependencies No The paper mentions PyTorch ('Py Torch: An Imperative Style, High-Performance Deep Learning Library. Co RR, abs/1912.0, 2019.') and notes that 'disentanglement_lib... is based on Tensorflow v1', but it does not specify version numbers for the PyTorch or other libraries used in their own implementation for reproducibility.
Experiment Setup Yes In our experiments, we set M = 1 and L = 100 since the MVHG and PL distributions are not concentrated around their mean very well, and more Monte Carlo samples thus lead to better approximations of the expectation terms. We further set β = 1 for MNIST and β = 0.1 for FMNIST, and otherwise γ = 1, and δ = 0.01 for all experiments.