Rényi Neural Processes

Authors: Xuesong Wang, He Zhao, Edwin V. Bonilla

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach across multiple benchmarks including regression and image inpainting tasks, and show significant performance improvements of RNPs in real-world problems. Our extensive experiments show consistently better loglikelihoods over state-of-the-art NP models.
Researcher Affiliation Academia 1CSIRO s Data61, Australia. Correspondence to: Xuesong Wang <EMAIL>.
Pseudocode Yes A.1. Pseudocode Algorithm 1 R enyi Neural Processes
Open Source Code Yes Our code is published at https://github.com/csiro-funml/renyineuralprocesses
Open Datasets Yes We evaluate the proposed method on multiple regression tasks: 1D regression [...] image inpainting [...] on three image datasets: MNIST, SVHN and Celeb A. [...] We also tested TND-D on the Extended MNIST dataset with 47 classes
Dataset Splits Yes The number of context points is randomly sampled M U(3, 50), and the number of target points is N U(3, 50 M) (Nguyen & Grover, 2022). We choose 100,000 functions for training, and sample another 3,000 functions for testing. [...] The number of context points for inpainting tasks is M U(3, 200) and the target point count is N U(3, 200 M). [...] We choose 20,000 functions for training, and sample another 1,000 functions for evaluation. [...] We use classes 0-10 for meta training and hold out classes 11-46 for meta testing under prior misspecification.
Hardware Specification No All the models can be trained using a single GPU with 16GB memory.
Software Dependencies No The paper does not explicitly mention any specific software dependencies with version numbers (e.g., programming languages, libraries, or frameworks).
Experiment Setup Yes We set α = 0.7 to train for VI-based RNPs and analogously α = 0.3 for ML-based baselines. [...] The number of samples K for the Monte Carlo is 32 for training and 50 for inference. [...] The input features were normalized to [ 2, 2]. [...] The input coordinates were normalized to [ 1, 1] and pixel intensities were rescaled to [ 0.5, 0.5]. [...] The noise level β is set as 0.3 for both training and testing.