Unsupervised Domain Adaptation (UDA) Project

A repository containing the Domain Adaptation notebook for the Deep Learning course of the University of Trento (AA 2021-2022).

Group Members

• Simone Alghisi
• Massimo Rizzoli

Proposed techniques

For this project we've used Discrimination Based Techniques for the most, by relying on the popular MMD loss for understanding the domain alignment. In general, we have implemented

• Deep Domain Confusion (between single and multiple layers), i.e. the classic implementation of MMD loss for solving UDA problems;
• DSAN (Deep Subdomain Adaptation Network), i.e. a more recent approach based on subdomain alignment;
• DRCN (Deep Reconstruction Classification Network), i.e. a different approach for UDA based on a multi-task approach using a decoder together with the classifier.

For more in depth details, please refer to the notebook.

Notes

Please keep in mind that the following project has been developed from Google Colab. To avoid possible incompatibilities issues, please load it and run it from there.