NeuroAI paper roundup #5: guide to NeurIPS

Winter is coming

I recently asked on X for suggestions for NeurIPS NeuroAI papers. You did not disappoint! I received several dozen replies referencing papers across the main conference and workshops. I realized I needed to be a bit more systematic about it. In brief:

• I fed all 3,600 papers from the main conference to GPT-4 turbo and extracted 197 NeuroAI and NeuroAI-adjacent papers.

• I compiled a list of the most relevant workshops.

How to survive NeurIPS

If you don’t want to get massively overwhelmed, I would suggest that you go through the list of main conference papers and focus on < 10 papers which are directly and highly relevant and timely to your work. Then, go see 2-3 workshops. Once the conference is past, you can draw down the papers throughout the year. Pace yourselves!

I don’t get a lot from being at the conference in person–between the 9AM start, the jetlag and bad coffee–and prefer to watch from the comfort of my own home at 1.5X with subtitles on. NeurIPS is one of the best conferences to watch from home IMO, as their tech is top-notch.

The main conference paper pile

I made a little visualization and interactive app in Airtable to allow you to browse all 197 NeuroAI papers from the main conference. You can also download the CSV if you want to do secondary analysis straight from the Airtable link.

It was a lot more papers than I anticipated! Our field is going strong! How I made this:

1. I went to the NeurIPS website and downloaded a json with all 3600 main conference papers, which I found using Cmd+Opt+I.

2. I fed all abstracts to GPT-4-Turbo to extract one of 5 broad categories of papers related to NeuroAI. I had 2 core categories, 1 neuro-data-analysis, 1 neuro-adjacent, and one “has nothing to do with NeuroAI” category. I re-used a script I made for the NeuroAI in Montreal conference, for which I was a co-organizer. Running this took 5 minutes and cost 30$ in OpenAI credits. I also asked GPT-4 for TL;DRs.

3. I manually went through the ~300 papers it selected in the top 3 categories to select the ones that I agreed were NeuroAI. I added in anything with the Neuroscience topic. I sorted the results such that anything anyone personally shared with me on X by Sunday night was shown at the start.

4. I embedded abstracts with a all-mpnet-v2 sentence transformer to find the nearest neighbours for each paper.

5. I dumped a CSV of everything and imported it into Airtable.

This took me about the time it would take me to read ~3-5 papers in detail. ALL the papers you shared with me on Twitter from the main conference were ALSO INDEPENDENTLY selected by GPT-4 and correctly bucketed into one of the 4 NeuroAI categories. These are truly the days of miracle and wonder.

The workshops

Here are the workshops that are neuro-heavy:

• UniReps: Unifying Representations in Neural Models

• Information-Theoretic Principles in Cognitive Systems (InfoCog)

• Associative Memory & Hopfield Networks in 2023

• Symmetry and Geometry in Neural Representations

These might contain some neuro:

• Gaze Meets ML

• XAI in Action: Past, Present, and Future Applications

• Generative AI and Biology (GenBio@NeurIPS2023)

Finally, don’t forget the NeuroAI social on Dec 13th (Wed).