Paper Note --- Transfer Learning via Dimensionality Reduction

Paper Background

Proceedings of the Twenty-Third AAAI COnference on Artificial Intelligence 2008
Author:

1. Sinno Jialin Pan
2. James T.Kwork
3. Qiang Yang

Hong Kong University of Science and Technology

---

This paper propose a new dimensionality reduction method to find a latent feature space, which minimize the distance between distribution of data in source domain and target domain in a latent space, thus we can use standard algorithms to train models.




From this formula we learned this kernel matrix K instead of learning the universal kernel k k <script type="math/tex" id="MathJax-Element-53">k</script>. However we need to ensure that learned kernel matrix does correspond to an universal kernel.

about kernel:
from Zhihu

Proved that learned kernel matix K is universal.

MVU 最大差异展开

---

Inspiration

For comparing two different distribution, we map them into a latent space by using kernel matrix, and compare the difference in such sapce will be more resonable.