Christian A. Naesseth

@naesseth
345 Followers
166 Following
67 Posts
Researcher interested in approximate inference, causality and artificial intelligence as well as their application to the sciences. Assistant professor in the Amsterdam Machine Learning Lab at the University of Amsterdam.
Websitehttps://naesseth.github.io/
Scholarhttps://scholar.google.com/citations?user=GQ6rOssAAAAJ
Twitterhttps://twitter.com/chris_naesseth
Show threadChristian A. NaessethMay 1, 2023
It turns out that when using Trajectory Balance to optimise the GFN it coincides with standard variational inference (VI) for the generative and reverse models! This connection allows us to leverage VI-improvements developed also for GFNs. See the paper for more details.
Show threadChristian A. NaessethMay 1, 2023
GFNs are generative models that uses auxiliary variables to recursively construct a distribution on the data space s_n to model a reward function R(s_n). For training, a reverse model of the generative models is constructed to enable optimisation.
Christian A. NaessethMay 1, 2023

New work by Heiko, Fredrik, Jan-Willem and myself interpreting Generative Flow Networks (GFN) as generative models trained by variational inference!

#TMLR #GenerativeAI #GFN #MachineLearning

https://openreview.net/forum?id=AZ4GobeSLq

A Variational Perspective on Generative Flow Networks

Generative flow networks (GFNs) are a class of probabilistic models for sequential sampling of composite objects, proportional to a target distribution that is defined in terms of an energy...

OpenReview
Christian A. NaessethFeb 28, 2023

Im very excited to announce that everyone's favourite Bayesian symposium is back for 2023!🚀🚀
The 5th Symposium on Advances in Approximate Bayesian Inference (AABI) will take place in 🏖️Honolulu Hawaii🌴, Sunday July 23rd, Co-Located with ICML!

Website: http://approximateinference.org
#aabi #machinelearning #bayes #icml

Approximate Inference

Christian A. NaessethNov 5, 2022

The insights of "Training Compute-Optimal Large Language Models" are interesting, but I wonder whether it still holds true if we consider multiple epochs? It seems reasonable that large models would be undertrained if we only see each datapoint a single time 🤔

https://arxiv.org/abs/2203.15556