Martingale Posterior Neural Processes

Authors: Hyungi Lee, Eunggu Yun, Giung Nam, Edwin Fong, Juho Lee

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

Reproducibility Variable Result LLM Response
Research Type Experimental The resulting model, which we name as Martingale Posterior Neural Process (MPNP), is demonstrated to outperform baselines on various tasks.
Researcher Affiliation Collaboration Hyungi Lee1, Eunggu Yun1, Giung Nam1, Edwin Fong2, Juho Lee1,3 1KAIST, 2Novo Nordisk, 3AITRICS
Pseudocode No The paper describes model architectures and training steps in narrative text and equations, but it does not contain formal pseudocode or algorithm blocks.
Open Source Code Yes We attached our code in supplementary material.
Open Datasets Yes MNIST We split MNIST (Le Cun et al., 1998) train dataset into train set with 50,000 samples and validation set with 10,000 samples. We use whole 10,000 samples in test dataset as test set.
Dataset Splits Yes MNIST We split MNIST (Le Cun et al., 1998) train dataset into train set with 50,000 samples and validation set with 10,000 samples. We use whole 10,000 samples in test dataset as test set.
Hardware Specification Yes We conducted all experiments on a single NVIDIA Ge Force RTX 3090 GPU, except for the image completion tasks presented in Section 5.2; we used 8 TPUv3 cores supported by TPU Research Cloud2 for the 2D image completion task.
Software Dependencies Yes Our codes used python libraries JAX (Bradbury et al., 2018), Flax (Heek et al., 2020) and Optax (Hessel et al., 2020).
Experiment Setup Yes For optimization, we used Adam (Kingma and Ba, 2015) optimizer with a cosine learning rate schedule. Unless specified, we selected the base learning rate from a grid of {5 10 4.50, 5 10 4.25, 5 10 4.00, 5 10 3.75, 5 10 3.50} based on validation task log-likelihood.