Machine learning for motion control using geometric algebra

Machine learning for motion control using geometric algebra
This studentship will be sponsored by ARM and carried out in the Signal Processing Group of the Cambridge University Engineering Department (CUED). The project will establish collaborations with researchers in the fields of bio-inspired robotics, control systems and inference and machine learning.

The aim will be to use Machine Learning to learn the geometric quantities associated with human motion. The mathematical framework of Geometric Algebra (GA) will be employed for this project. In particular we are interested in studying the application of a 5D representation of 3D Euclidean space, called Conformal Geometric Algebra (CGA). In this algebra primitive objects (blades) can represent points, lines, planes, circles, spheres and volume elements. These objects all inherit a multiplication rule from the underlying CGA. Transformations between these objects (rotations, translations and dilations) are also represented as simple objects in this algebra. It is these primitives of GA that we aim to learn.

As well as understanding the mathematical framework and programming the machine learning environment, the project will also involve instantiating aspects of the machine learning algorithms in hardware. There is a substantial literature of GA implementations on FPGA hardware, mainly for blade multiplication, and here we will investigate the feasibility of extending this work to more complex machine learning workloads.