Advances in Bayesian Inference for Model Generalization
Sanae Lotfi
Details
In this talk we'll hear from the lead author behind award winning ML research presented at the 2022 ICML conference.
The topic is the main insights behind the following paper: https://arxiv.org/pdf/2202.11678.pdf
The topic is the use of Bayesian Inference and model generalization. Bayesian inference has long been used to select model parameters for ML problems, both as an alternative and compliment to traditional cross validation.
This paper reveals how standard approaches for Bayesian approaches using the marginal likelihood can have unexpected pathologies for model choice, and presents a modified approach that offers better results.
The conclusions are rooted in both theoretical insights and experimental results, and provide an enlightened way to think about how optimization techniques can and cannot lead to generalized out-of-sample results.
Developing ML models that generalize well to unseen data is a critical area of expertise in data science, and this talk reveals new approaches to thinking about this problem.
While this event is FREE, tickets are required & space is limited!
Attend this event
.svg)