Neural Information Processing Systems (NeurIPS) is one of the top machine
learning conferences in the world and is the venue where seminal works such as
AlexNet (Krizhevsky et al 2012) and transformers (Vaswani et al 2017) were
first disclosed. We are excited to announce that Blueteam AI's research
scientists will be publishing and presenting at the 2023 NeurIPS conference in
New Orleans, LA and we look forward to meeting you in person.
Our staff will be presenting the following paper:
Liang, Feynman T., Liam Hodgkinson, and Michael W. Mahoney. "A heavy-tailed algebra for probabilistic programming." Advances in Neural Information Processing Systems 36 (2024).
This work addresses modelling heavy tailed data where outliers ("black-swan events")
are more likely as commonly encountered in insurance claims, financial market shocks,
and extreme weather events. Whereas traditional methods break down due to the extreme
values introduced by outliers, in this work we propose a classification into heavy-tailed
families and derive closed-form analytic solutions for how these families behave under
common probabilistic operations. The result can be used "quick and dirty"
estimation of tails of distributions after transformation as well as to inform
better initialization of heavy tailed distribution approximators.
For those interested in more details, please see our
OpenReview, arXiv,
and the following short video explaining more: