How to visualize modern AI research
A tool for visualizing AI research at NeurIPS this year is more important than ever as the conference sets a new record with more than 4,500 accepted publications.
A tool for visualizing AI research at NeurIPS this year is more important than ever as the conference sets a new record with more than 4,500 accepted publications.
Back in 2020, IBM researcher Hendrik Strobelt built a visualization tool to make it easier to find and present papers at NeurIPS. The world’s top AI conference was held virtually that year because of the pandemic, and though in-person meetings are back, the visualization tool has become a permanent feature.
You can use it to search this year’s 4,543 accepted papers by author and title. You can also see how the papers relate to each other thematically, clustered into clouds labeled “language models”, “multimodal LLM”, and “time series,” among others. In collaboration with another IBM researcher, Benjamin Hoover, Strobelt built the tool with a large language model pre-trained on 300,000 AI-related abstracts on ArXiv, using UMAP to contextualize the embeddings in relation to each other.
They used the tool in 2021 to look back over 60 years of AI research and pick out the 3,300 most influential papers by year and number of citations. One trend that jumped out was the shift from classical machine learning to deep neural networks, which underlie modern AI. A similar trend can be seen at NeurIPS, short for Neural Information Processing Systems, where the number of papers accepted was 50 times greater this year than in 1987 when the conference started.