Key information:
| Student | Alexander Lowe |
| Academic Supervisors | Alexander Serb, Wendy Hall, Themis Prodromakis |
| Cohort | 2 |
| Pure Link | Active Project |
We want to produce a model which can perform fluid reasoning, to have the ability to logically solve problems it has not necessarily seen before. A sufficiently large statistical learning model can imitate this behaviour, by just learning all the possible answers ahead of time, however this becomes infeasible with scaling and so a different approach is required. Statistical methods provide classifications and pattern recognition in datasets, information which can be packaged into higher level variables . These are the symbols in a symbolic AI. Symbols can be manipulated at the required higher level of abstraction, such as to allow fluid reasoning, by operations prescribed by so-called cognitive architectures. Several such architectures have been proposed in literature. The question then is, how best can these systems be translated into hardware?
By studying the underlying mathematical machinery of cognitive architectures we seek to systematically determine the properties that result in hardware-friendly architectures. Understanding key features such as bitwise interactions throughout an operation or the significance of the choice of representations in the system allows us to choose metrics by which to judge the proposed architectures. With these metrics in hand we can build systems to automate surveying the space of symbolic operations available and so determine the most hardware-efficient symbolic architecture implementations.
This work sits as part of a group of projects striving towards a common goal, the design and construction of a hardware accelerator dedicated to performing symbolic processing operations. The results of this work will inform the design process of the accelerator through an understanding of the fundamental mathematics underpinning energy-optimised symbolic AI hardware.