Source Model Selection for Transfer Learning of Image Classification using Supervised Contrastive Loss

Transfer learning is a framework that improves performance of target task by transferring knowledge from training source task. As deep learning research accumulate, more source models can be easily obtained. In time series domain, the Mean Silhouette Coefficient of the set of feature vectors which forward propagated through source models is used to select the best source model performs target task. But for image classification, the model which is better at generalization could have the lower coefficient. To adjust this, we propose to use another measure, Supervised Contrastive Loss. In this work, we evaluate which measure is better to select the best model. We present the superiority of using the supervised contrastive loss through the comparative experiment.