Caption: Wordcloud of all NeurIPS 2023 titles
Welcome to the hub for all things NeurIPS 2023! We scraped the data for all 3500+ NeurIPS projects and dove into the depths of Hugging Face, GitHub, LinkedIn, and Arxiv to pick out the most interesting content.
In this repo, you will find:
- Data Analysis: detailed analysis of the titles and abstracts from NeurIPS 2023 accepted papers
- Awesome Projects: synthesized collection of 40 NeurIPS 2023 papers you won't want to miss
- Conference Schedule: comprehensive listing of all NeurIPS 2023 projects (title, authors, abstract) organized by poster session and sorted alphabetically
The raw data is included in this repo. If you have ideas for other interesting analyses, feel free to create an issue or submit a PR!
For now, insights are organized into the following categories:
- Authors
- Titles
- Abstracts
🔍 For the data analysis itself, check out the Jupyter Notebook!
🔍 And check out the blog post synthesizing the results here.
The top 10 authors with the most papers at NeurIPS 2023 are:
- Bo Li: 15 papers
- Ludwig Schmidt: 14 papers
- Bo Han: 13 papers
- Mihaela van der Schaar: 13 papers
- Hao Wang: 12 papers
- Dacheng Tao: 11 papers
- Bernhard Schölkopf: 11 papers
- Masashi Sugiyama: 11 papers
- Andreas Krause: 11 papers
- Tongliang Liu: 11 papers
There were 13,012 unique authors at NeurIPS 2023, up from 9913 at NeurIPS 2022.
This continues the exponential explosion of unique authors over the past decade.
- The average number of authors per paper was 4.98, up from 4.66 at NeurIPS 2022.
- Additionally, there were a handful of single-author papers, in contrast to NeurIPS 2022, where the minimum number of authors was 2.
- The paper with the most authors was ClimSim: A large multi-scale dataset for hybrid physics-ML climate emulation
- The average title length was 8.72 words, up from 8.48 at NeurIPS 2022. This continues an ongoing trend of title lengthening:
22% of titles introduced an acronym, up from 18% at NeurIPS 2022.
- 1.3% of titles contained LaTeX, whereas none of the titles at NeurIPS 2022 contained LaTeX.
- The longest abstract was from [Re] On the Reproducibility of FairCal: Fairness Calibration for Face Verification, which has 373 words.
- The shortest abstract was from Improved Convergence in High Probability of Clipped Gradient Methods with Heavy Tailed Noise, which has 29 words.
- Out of the 3581 abstracts, 675 explicitly mention GitHub, including a link to their code, models, or data.
- Only 79 abstracts include a URL that is not GitHub.
Using a CLIP model, we zero-shot
classified/predicted the modality of focus for each paper based on its abstract.
The categories were ["vision", "text", "audio", "tabular", "time series", "multimodal"]
.
By far the biggest category was multimodal, with a count of 1296. However, the
CLIP model's inclination towards "multimodal" may be somewhat biased by trying
to partially fit other modalities — the words multi-modal
and multimodal
only
show up in 156 abstracts, and phrases like vision-language
and text-to-image
only appear a handful of times across the dataset.
Themes occurring frequently include:
- "benchmark": 730
- ("generation", "generate"): 681
- ("efficient", "efficiency"): 963
- "agent": 280
- ("llm", "large language model"): 238