Generative Marginalization Models

Authors: Sulin Liu, Peter Ramadge, Ryan P Adams

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate marginalization models (MAM) on both MLE and EB settings for discrete problems including images, text, molecules and phyiscal systems. We compare MAMs with baselines that support arbitrary marginal inference3: Anyorder ARM [24], ARM [38], Parallel Any-order ARMs (P-AO-ARM) [24] and Probabilistic Circuit (PC) [54].
Researcher Affiliation Academia 1Princeton University. Correspondence to: Sulin Liu <sulinl@princeton.edu>.
Pseudocode Yes We present the algorithms for training MAM for maximum likelihood and energy-based training settings in Algorithm 1 and Algorithm 2.
Open Source Code Yes Code is available at github.com/Princeton LIPS/Ma M.
Open Datasets Yes We evaluate MAMs on Binary MNIST [63], CIFAR-10 [37] and Image Net32 [14, 10]. We evaluate MAM on the molecular generation benchmark MOSES [55] refined from the ZINC database [72]. We train a character-level generative model on Text8 [45].
Dataset Splits Yes We use the training and test split of [63] provided in https://github.com/yburda/iwae/tree/master [6]. The CIFAR-10 dataset [37] comprises 60,000 32x32 color images across 10 classes, split into 50,000 training and 10,000 test images.
Hardware Specification Yes All models are trained on a single NVIDIA A100. The evaluation time is tested on an NVIDIA GTX 1080Ti.
Software Dependencies No The paper mentions software components like 'Adam' and 'Adam W' for optimization, and 'U-Net' and 'Transformer' for neural network architectures. However, it does not specify version numbers for any key software libraries (e.g., 'PyTorch 1.x' or 'TensorFlow 2.x'), programming languages, or specific solvers used to run the experiments.
Experiment Setup Yes Batch size is 128 for MNIST and 32 for CIFAR-10 and Image Net. Adam is used with learning rate 0.0001. Gradient clipping is set to 100. Both AO-ARM and MAM conditionals are trained for 100 epochs on MNIST, 800 epochs on CIFAR-10, 16 epochs on Image Net.