Freiburg University, Hildesheim University and other universities in Germany jointly | zero shot automl with pre trained models (zero sample automl based on pre training model)

author ：Ekrem Ozturk,Fabio Ferreira,Hadi S. Jomaa etc.

brief introduction ： Given a new data set D And a low calculation budget , How should we choose a pre training model to fine tune to D, And set fine-tuning superparameters without the risk of over fitting , Especially in D Very small case ？ ad locum , The authors extend automated machine learning (AutoML) To make the best of these choices . author A zero sample agent model is learned by the domain independent meta learning method , The model allows for new data sets when testing D Choose the right deep learning (DL) The Conduit （ It includes pre training model and fine tuning super parameters ） Just give a description D Trivial meta features of , For example, image resolution or class number . To train this zero sample model , The author collects many on a large data set DL Performance data of pipeline , And perform meta training on these data , To minimize pairwise sorting objectives . For advancement ChaLearn AutoDL Challenge benchmark , The author evaluates the author's method under strict time constraints , Obviously superior to all competitors who challenge .

Paper download ：https://arxiv.org/pdf/2206.08476.pdf