This repository contains codes of our paper Transferability vs. Discriminability: Joint Probability Distribution Adaptation submitted to ECAI 2020. This paper proposes a novel and frustratingly easy Joint Probability Distribution Adaptation (JPDA) approach, to replace the frequently-used joint maximum mean discrepancy metric in transfer learning. During the distribution adaptation, JPDA improves the transferability between the source and the target domains by minimizing the joint probability discrepancy of the corresponding class, and also increases the discriminability between different classes by maximizing their joint probability discrepancy. Experiments on six image classification datasets demonstrated that JPDA outperforms several state-of-the-art metric-based transfer learning approaches.
The average accuracies on the PIE dataset are shown in Table 2. JPDA outperforms all the joint MMD based approaches in most tasks, and achieve an accuracy improvement of 4.38% compared with JDA.
The code is MATLAB code works in Windows 10 system.
Code files introduction:
demo_classify_other.m -- demo file, joint probability distribution adaptation (JPDA) over 4 cross-domain image classification tasks on datasets COIL20, USPS and MNIST.
demo_classify_office.m -- demo file, JPDA on 12 cross-domain image classification tasks on dataset Office Caltech-256.
demo_classify_pie.m -- demo file, JPDA on 20 cross-domain image classification tasks on dataset PIE.
JPDA.m -- function file, it's the implementation of JPDA approach. Please find the specific input/output instructions in the function comments.
Please find the details and references described in our paper.