Learning Curve Extrapolation with PFN’s

Tom Viering

Assistant Professor at Tu Delft

Abstract: Machine learning usually requires large amounts of data to work well. But how much data is exactly required for sufficient performance? This can be estimated using learning curves. Learning curves plot a machine learning model’s performance versus the amount of data it is trained on. Note: the difference with performance versus epoch curves, which are usually referred to as learning curves in the neural net community. By running a pilot study with a small amount of data, we can make a small learning curve. The learning curve can be extrapolated to obtain a rough estimate of the data requirements for a particular performance. To this end it is important to know how to best extrapolate the learning curve. In this project we will investigate how to do this using PFN’s.

Keywords: Learning Curve, PFN, How much data
Scientific area: Artificial Intelligence, Machine Learning, Fundamentals
Visiting period: 1st of May – end of June
Visiting Lab: Frank Hutter’s Lab in Freiburg

Bio: I graduated my PhD in May 2023. The topic of my PhD was on safety on machine learning and since 2024 I am an assistant professor in the research group PRB of TU Delft. My PhD thesis has highlighted that learning curves, which plot performance versus amount of training data, can show surprising behaviors. For example, we have highlighted several examples where more data can lead to worse performance. Ever since, I have been interested in characterizing learning curve behavior in more detail. In particular, I believe learning curves can be used to answer the age-old question: how much data is needed? By extrapolating learning curves we I believe we can answer this question.