Poster Session
in
Workshop: Scientific Methods for Understanding Neural Networks
"Twin Studies" of Factors in OOD Generalization
Victoria R. Li · Jenny Kaufmann · David Alvarez-Melis · Naomi Saphra
Abstract:
Transformer models trained on an ambiguous classification task demonstrate diverse out-of-distribution behavior with different hyperparameter and random seed settings. We analyze this model population across and at the end of training, describing how we can leverage such natural variation to draw conclusions about determinants of model performance and generalization.
Chat is not available.