||The capture, reconstruction and synthesis of facial expressions often involves specialized hardware support and considerable computation time. This prohibits its widespread deployment and use in real-time applications. In this paper, we aim at tackling this limitation via a learning-based approach, which is efficient and requires only modest hardware support. Our approach is based on a semi-supervised manifold alignment framework, where feature points extracted from 2D face images are aligned with data expressed as morph-target values for a 3D face model. By applying a kernel embedding method known as kernel locality preserving projections (KLPP) and a method for solving the pre-image problem in kernel methods, our framework is capable of handling nonlinearity and is defined everywhere. Experiments are conducted to demonstrate two possible applications of our proposed framework: 3D reconstruction of facial expressions and dynamic synthesis of facial expression sequences.