based on 60 years demo from Hendrik Strobelt and Benjamin Hoover from MIT-IBM Watson
AI Lab;
data from Lee Campbell; tested by Marc'Aurelio Ranzato
The NeurIPS conference has been around for more than 35 years, and interest in the fields of AI/ML is still rapidly growing. A diversification of interests has birthed many sub-fields within the fields, making it harder for novices and senior researchers alike to orient themselves and their work within the historical context of research published at NeurIPS. We created an interactive demo to investigate the papers from the last 35+ years. The demo not only reflects on the past, but it also allows users to position abstracts of their own novel ideas into the research landscape carved by the last decades of NeurIPS papers.
The tool provides many options for discovery, like:
To figure out the neighborhood in which your next idea would fit, you can run your abstract through the same pipeline as all other papers and position your abstract with the PIN function.
Simply copy and paste your abstract into the field and hit "place pin".. it might take some seconds but an orange dot will soon appear indicating the position of your paper in the landscape.
Caution: if you use the "reduce overlap" function, the request might take 2-3 times longer.
By dragging your mouse over the field of papers, you can quickly define a cluster of interest. After releasing the mouse button, all papers under the selection rectangle are summarized by (1) most common single words and (2) most common bi-grams of paper titles. We also show the yearly trend of paper count in your cluster.
If you think someone else should see what you just discovered, you can share your cluster selection by the "copy link" or "tweet" buttons.
Search for substrings in author names or paper titles by typing in the search bar and pressing the button "author" or "title". To remove the search filter just hit "clear".
Each paper is colored by its year of publication (green to yellow). You can select papers either by a 5-year period or by using the yearly filter that shows all papers up until your selected year. See below all papers from 2020 or younger:
You can select to (not) show some default clusters that we think might be interesting. You can also show the paper dots such that they do not overlap. This guarantees that you can hover over each single paper but requires more space and compute to calculate the overlap removal.
Here is an example of dots arrangement with overlaps removed:
By inspecting the landscape I found this little island in the upper left region of the map and selected it. Immediately I can see that the cluster summary shows me terms like "fairness" or "fair classification". I also see that the Papers Per Year histogram shows that the field is relatively new (only bars past 2015). Because I think that is interesting, I copy the link by the "copy link" button and send my colleagues who work on fairness an email with the link.
We wrote a paper published at another conference on explainable AI for Seq2Seq models and I wonder where this would live in the landscape of NeurIPS papers. So I paste the abstract into the "abstract" field and hit "place pin". The orange dot appears in the language cluster, which is confirmatory. There are some older papers in its surrounding area, perhaps these can inspire new ideas.
From animating the yearly filter I can see how the field has developed over the years — from early papers published on Neuroscience, to classical NN and ML approaches, to modern Deep Learning in it's many fashions in recent years. In this small way I too can experience the "explosion" of the field of AI.