The SPIKE group has a strong research track record on symbolic knowledge-driven, model-based AI. Although, pure data-driven machine learning have proven to be effective in solving circumscribed tasks, exceeding human performance, it lacks generalisability, robustness to out of distribution data, and interpretability. In SPIKE, we explore and develop novel archtectures for combining advanced symbolic AI solutions with machine learning in order to overcome their respective limitations and leverage upon their renown advantages.

Neuro-Symbolic Reinforcement Learning

Despite many successes in mastering complex tasks, reaching in some cases better performance than humans, pure RL algorithms truggle to discover and exploit teh structure of a task and relevant abstractions. We explore Neuro-Symbolic RL approaches that makes use of state-of-the-art symbolic machine leanring systems to learn state abstractions, abstraction herachcal stricture of tasks and interleave this learning togetther with the RL exploration. This with the objective of improving generalisability and transferbaility of RL policies, whilst facilitating task decomposition and modularity.

For a list of our published work on this topic, please visit our publications website.

Neuro-symbolic Learning

We have developed both sequential and end-to-end methods for learning interpretabel knowedge from raw (unstructured) data. In the squential method, pre-trained networks can be used to extract features from raw data, which in turn can be used to learning general knowledge needed to solve a given task. The robustness of the symbolic learning mechanisms used in our research enables such such metho to be used also on data that are outside th distributed used to traind the network. In the end-to-end method, we address the challenge of training neural component and symbolic learning simulatenously supervised only by the signal provided by dowstrem labels. No ground truth information is given about latent concepts that the neural compoenent has to learn.

For a list of our published work on this topic, please visit our publications website.

Neuro-symbolic Reasoning

Neuro-symbolic reasoning approaches tend to either combine a neural perception component with a symbolic reasoning component, or performing symbolic reasoning in continuous vector spaces rather than discrete space. The ultimate objective be to provide neural copmonents with symbolic reasoning capabilities that can help them improve their interpretability and enable generalization beyond the training tasks. We have been exploring both streams of neuro-symbolic reasoning by (i) tackling the challenges of capturing common-sense reasoning in continous space and (ii) improving the scalability of neuro-symbolic reasoning tasks facilitating fast training of latent space from raw data and using symbolic reasoning to give seantic meaning to the discovered latent features by means of symbolic optimisation within the scope of existing prior knowledge.

For a list of our published work on this topic, please visit our publications website.