Our paper on providing reliable neuro-symbolic classification by combining search and learning based on conformal prediction has been accepted at NeurIPS 2026. I will be presented the poster in the Paris edition of this NeurIPS.
For those are not up-to-speed: NeurIPS is one of the world’s premier academic conferences dedicated to presenting groundbreaking research in artificial intelligence, machine learning, and computational neuroscience. It has grown so large (and the global atmosphere around travel, visas, CO2, costs etc. has grown to toxic) that the 2026 edition will be hosted at three official locations simultaneously for the first time. The locations are Sydney, Atlanta and Paris and each come with its own keynotes, main track orals, poster sessions, tutorial and workshops. What a time to be alive!
The story of this paper is a special one. The seed idea for this paper appeared in the context of a group outing in the first two months or so that I joined the VU’s learning and reasoning group. I had been looking at conformal prediction for uncertainty quantification and wondered whether symbolic knowledge would allow for generating better uncertainty quantification. At the time (2020) there were no such methods available, although in hindsight we can infer some work on this general direction had probably already started by then.
The idea of the offsite group outing was to come up with a(n idea for a) paper in two days. This turned out to be too ambitious. Our paper only appeared roughly three years after its initial conception during the outing, and went through many rounds of refinement in between in an on-and-off fashion.
Besides all the learning on conformal prediction, search over graphs, combinatorics, family-wise error rate control etc. there were some lessons in this paper for me on process:
- sometimes you just have to start working on an initial idea and iterate on it
- getting into a new topic and delivering something useful there takes a lot of grit.
- do not disregard your reviewers’ concerns (note: this was mostly pre-LLM reviewing)
- hang in there, and try to take the viewpoint of continual learning