The main idea behind transfer learning is enhancing inter-domain transferability and intra-domain discriminability. Ben-David’s Theorem demystifies transfer learning and fascinates me greatly. I wrote survey of UDA.pdf as a brief summary for my research experience on Unsupervised Domain Adaptation.
During my research on traditional Unsupervised Domain Adaptation, I wanna find a cross-atlas diagnose solution for Alzheimer's and CDLS inspired me. I derived a multi-site cross-atlas diagnose model from CDLS and reformulated it into a QP form in multi-source domain adaptation. This model failed in my data, and actually the multi-site cross-atlas diagnose scenario is not pratical.
I met a lot of Transformer models during intern and VectorNet is the most elegant and inspiring work among them. I wrote self-attention and GCN.pdf for a group meeting on how to model interactions among traffic agents.