NeurIPS 2021 Introduces New Datasets and Benchmarks Track
The NeurIPS 2021 conference launches a dedicated track for exceptional datasets and benchmarks, accepting 163 submissions. This new initiative aims to elevate data-oriented work and improve dataset development practices across the machine learning community.
The NeurIPS 2021 conference introduces a new Datasets and Benchmarks track to highlight exceptional work in creating high-quality datasets and insightful benchmarks. This initiative serves as a dedicated venue for researchers to share data-oriented contributions and discuss ways to improve dataset development across the machine learning field. The conference organizers express great excitement about the quality and potential impact of the accepted work.
The track accepts 163 submissions that cover a wide array of innovative machine learning applications. Notable projects include a dataset of 3D garments with sewing patterns, an open-source Mandarin speech dataset featuring eight subdialects, and an MRI reconstruction dataset designed for quantitative clinical evaluation. Other accepted papers explore end-to-end document understanding, Bayes error estimation, and self-supervised pretraining without human labeling.
Beyond technical advancements, several accepted papers address critical ethical considerations in machine learning data creation. Research projects investigate the mitigation of dataset harms through proper data stewardship and construct visual datasets to study the ongoing effects of spatial apartheid in South Africa. These diverse submissions demonstrate that the new track successfully captures both cutting-edge technical benchmarks and essential socio-technical discussions.