This project is provided as a quick and easy programing evidence for the results presented in the this paper which addresses Transfer Learning problem.
In this project we show that when the model trained in the source domain (MNIST dataset) is directly applied to the target domain (MNIST-M dataset) it could only achieve 0.49 accuracy. Fortunately, when the method proposed of the paper is employed for Transfer Learning, specially when the Temperature parameter is involved, the accuracy of the prediction in the target domain increased to 0.68.
The project is developed based on Python 2. In this project, PyDeep package is utilized for training and sampling form RBMs. Also TensorFlow version 1.x is used for neural network training.
Please feel free to contact me if you have any comment or question regarding
the project or the paper.
Email address: h.sh.farahani@gmail.com