Ask Oxford Professors: The Interplay Between Machine Learning and Semantic Reasoning

Artificial intelligence (AI) is a widely used term that conjures notions of fantasy, the future, or even threat. This is not surprising considering the multitude of movies which dramatise the role of artificial intelligence and what it may become.

But what is Artificial Intelligence?

In reality, artificial intelligence is a branch of computer science which aims to “understand and build intelligent entities by automating human intellectual tasks”.

These processes have contributed to numerous technological advances across various industries, for example. self-driving cars, technology for diagnosing cancer, revealing fraud in financial services and new data processing techniques. It is now quite common to see articles about the latest AI development — check out these robots which flip burgers!

Artificial intelligence has two main strands: Machine Learning and Semantic Reasoning.

Machine Learning is a process of learning from experience. There are three types of machine learning, supervised, unsupervised and reinforcement learning. When a machine learning model makes decisions or predictions using past or labelled data, this is supervised learning. Whilst in unsupervised learning, the goal is to identify inconsistencies, patterns or relationships within a set of input data. Reinforcement learning uses rewards, such as positive or negative feedback to train the model. For those interested in Psychology, the latter reminds me of behavioural studies, such as Skinner’s 1938 paper on Operant Conditioning.

Machine Learning models are used in a wide range of applications where the purpose is to model an output with data. For example, machine learning models are used to predict basket spend on an e-commerce website, credit scores, fraudulent transactions etc. Unsupervised models can be used to cluster customers for marketing purposes.

Semantic Reasoning on the other hand, is the ability of a system “to make logical deductions from the information that is explicitly available”. RDFox is a knowledge graph and semantic reasoning engine, which can derive new data from the data that it is explicitly given, using ontologies and rules.

Knowledge Graphs are already used in a vast number of applications, however the ability to carry out correct and fast reasoning has been limited across the industry. RDFox prides itself on being the first high-performance knowledge graph built from the ground up with semantic reasoning in mind, allowing for correct and fast reasoning, at scale and on the fly.

Can Machine Learning and Semantic Reasoning complement each other?

There is speculation over whether these two strands of AI will work together or in opposition as the industry develops. To elaborate on the debate, we’ve asked three leading experts on Artificial Intelligence, Professors from the Computer Science department at the University of Oxford, their thoughts on the matter.

Professor Ian Horrocks

“People are working on this problem and how to integrate machine learning and semantic reasoning to try and get the best of both worlds. In the short term, approaches are already being used at a simple level. For example, using machine learning approaches to improve knowledge creation and curation.

One of the big problems with Knowledge Graphs, is where does the Knowledge come from? It can be difficult to capture knowledge and time consuming. So we are doing a lot of work with machine learning to try and improve both the capturing of knowledge and curating it into a knowledge graph. The aim of this is to improve the quality and it provides a really nice example of how the two types of AI can really work together.”

Professor Bernardo Cuenca Grau

“I think the two technologies are compatible in several ways. Like Ian mentioned, you could use machine learning to improve the quality of knowledge graphs. For example, there are new techniques for suggesting new connections in knowledge graphs which use machine learning models. There are techniques for integrating different knowledge graphs and finding corresponces between nodes which are based on machine learning as well.

But you can also use knowledge graphs to improve machine learning. For example, as a platform for integrating data from multiple sources, providing data which is more relevant for machine learning algorithms, and to make machine learning models more explainable and transparent.”

Professor Ian Horrocks

“I think there is a lot to be done to make the integration of the two technologies much tighter. As Bernardo has indicated, machine learning developers are interested in using knowledge-based systems to improve explanation, so that they can understand what the algorithms are doing. By using semantic reasoning you can have a much clearer explanation of what is going on rather than just having a black box that seems to be doing what you want it to, but it is difficult to tell exactly what is going on.”

Professor Boris Motik

“I agree, there could be great benefits from using semantic reasoning to explain machine learning. For example, you could use semantic reasoning to prove properties about that model, so that you can get some hard and fast guarantees, rather than just an ‘it works’ statement. Underneath this is quite a challenging theoretical question, on the surface, these two things seem very different, which is why we don’t have any good solutions yet. But sooner or later it is going to be solved”

Professor Bernardo Cuenca Grau

“Sometimes machine learning and semantic reasoning might be viewed in competition. For instance, you could tackle fraud detection with machine learning by training a classifier or machine learning model using historical data on transactions and past fraudulent action, so when a new transaction is made the model can predict if the transaction is fraud or not. It does this essentially by learning patterns in the transactions. These predictions can be very good, but going back to an earlier point, the lack of explainability is an issue, as someone’s card might get blocked for seemingly no reason.

You could also approach this problem with semantic reasoning by speaking to a fraud specialist, deducting knowledge on what constitutes fraud, for example: is fraud linked to particular servers? Are there multiple connections at the same time? Then this knowledge can be encoded with rules. The system would then flag potential fraud and can provide information on why that action was flagged. Although this might require extra work in terms of encoding knowledge, you end up with an explainable and transparent system.

Both methods work, but they are different. There are a lot of problems which could benefit from using both strands of AI and combining them in a smart way.”

Professor Boris Motik

“I firmly believe this is going to be the next, big breakthrough in AI. We need to look at a deep level, into deeply interconnecting machine learning and semantic reasoning.”

‍

It appears that there is great scope for the integration of machine learning and semantic reasoning and this perhaps might hold the answer for the next big advancement in the artificial intelligence industry. We look forward to seeing the applications which will harness these technologies and use their relative strengths.

If you have a machine learning product that you think would work well with RDFox, get in touch by emailing info@oxfordsemantic.tech.

Request a free 30-day trial license now!

‍

Enjoyed this article? You might also like ChatGPT’s ‘Snow White’ Problem: The Danger of Common Knowledge.