Neural-symbolic learning of interpretable high-level knowledge from raw data

The main aim of this project is to develop a framework that combines neural networks with a symbolic layer. The framework should be able to learn knowledge in terms of interpretable symbols and rules from raw data, without them being hard-coded.

The key research objectives/research questions are:

- Learning of symbolic knowledge from real-world data, as opposed to well-structured and noiseless data

- Using answer set programming (ASP) or similar paradigms to represent knowledge

- Incorporating a neural network as the perception layer and enhancing it with a symbolic reasoning layer

- Learning symbolic latent concepts that are used for solving a downstream task but are not directly available as labels

- Minimising the use of hard-coded knowledge

- Creating an explainable framework, in which the learned knowledge is human-readable and understandable

- Applying the framework to real-world data, such as text, handwritten digits or more complex image

To achieve these objectives, I will:

- Enhance existing frameworks that combine neural networks with symbolic AI, such as Embed2Sym or NSIL

- Introduce signals for latent concepts that narrow down the hypothesis search space

- Incorporate methods to guide the learning of the symbols, such as active learning, interactive machine learning or reinforcement learning from human feedback

- Develop processes to make the symbols and rules readable, e.g. using natural language processing

The expected novel contributions are:

- Learning of ASP hypotheses or other forms of logic-based knowledge from data with fewer hard-coded concepts than the current state of the art problem domains that are more challenging than what current papers use

- Combining symbolic AI with human-in-the-loop methods

- Explainable decision-making of a neural-symbolic framework

This PhD is relevant to the following EPSRC research area(s):

- Artificial intelligence technologies

- Logic and combinatorics