This document discusses various transfer learning techniques for machine learning, including domain adaptation and small sample learning. It proposes three methods for unsupervised domain adaptation that use graph or hypergraph matching to minimize domain discrepancy: 1) Graph Matching, 2) Hypergraph Matching, and 3) Graph Matching with representation learning. For small sample learning, it discusses approaches for few-shot learning and zero-shot learning, and proposes a two-stage solution for few-shot learning that learns a discriminative low-dimensional space and estimates class variance, and a method for zero-shot learning that matches features to semantics. Evaluation on standard datasets shows the proposed methods achieve competitive performance.