This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
Interpretability 2.0
Accepted at the 
Interpretability 2.0
 research sprint on 
May 10, 2023

Detecting Phase Transitions

Summary Our aim was to develop tools that could detect phase transitions (parts of training in which the model quickly learns a particular subtask) purely from weights. We ended up blocked by finding suitable datasets in which to study phase transitions. We attempted several techniques to control and induce transitions: "graduating the data" and studying bounded polynomials of varying difficulty, but these all ran into problems. We also looked at well-known tasks with transitions (grokking) and learning without transitions (MNIST & CIFAR-10). We hereby lay the seeds for (future) phase detectors. You can find a GitHub repo with the (ongoing) work here. Notebooks: (graduated) MNIST, bounded polynomials, CIFAR-10.

Jesse Hoogland, Lucas Texeira, Benjamin Gerraty, Rumi Salazar, Samuel Knoche
4th place
3rd place
2nd place
1st place
 by peer review