Roundtrip: UCI and Image Benchmarks for Density Estimation

Datasets from the 'Roundtrip' paper provide a collection of UCI machine learning benchmarks and standard image datasets for evaluating deep generative models. The collection includes tabular data from domains such as activity recognition, protein structure, finance, audio, and particle physics, alongside the MNIST and CIFAR-10 image datasets. It is intended for benchmarking density estimation and generative modeling techniques across diverse data types.

Use Cases

• [Benchmarking density estimation models] based on [tabular UCI datasets like AReM, CASP, and BANK]
• [Evaluating generative model performance on image data] based on [inclusion of MNIST and CIFAR-10 datasets]
• [Testing outlier detection algorithms] based on [the ODDS datasets Shuttle, Mammography, and ForestCover]
• [Training multimodal neural networks] based on [the mix of tabular, audio, and image data formats]

Strengths

• [Cross-platform presence] indicates the dataset is established and used in published research.
• [Provides specific data sizes] for several constituent datasets, such as YPMSD with 515,345 rows and 90 columns.
• [Includes a variety of data types] covering tabular, audio, and image domains for broad benchmarking.

Limitations

• [Metadata is incomplete] as column names, total row counts, and specific update dates are not provided across sources.
• [Dataset is a composite collection] of pre-existing benchmarks, offering no new primary data collection.
• [Source descriptions are repetitive] across the provided platform entries, limiting independent verification of details.

Provenance

Source

Compiled from UCI Machine Learning Repository and standard image datasets (MNIST, CIFAR-10) for the 'Roundtrip' research paper.

Collection Method

Likely aggregated from public repositories for use as a unified benchmark suite.