Artificial Intelligence (AI) has transformed many areas of life and business. Its impact has been hard to quantify. Equally, it is increasingly common to see examples of where AI and machine learning have provided wrong answers. AI expert Amy Hodler explores how AI algorithms can go awry and highlights how graph technology can help

Amazon facial recognition technology mistakenly identified three-time Super Bowl champion Duron Harmon of the New England Patriots as a criminal. Microsoft’s AI Chatbot was trolled by users and provoked into repeating offensive phrases. The impact of mistakes like these can be enormously damaging to a brand. Organisations are keen to address the problem. The issue can be distilled into three elements of a new concept, ‘explainability’.

Explaining explainability

The concept of explainability encompasses data, predictions, and algorithms. What data was used to train the model, and why? What features and weights were used for this particular prediction? And what are the individual layers and the thresholds used for a prediction?

An organisation could have the best AI system in the world, but if underlying data has been manipulated or is incomplete, they cannot rely on that system. It is imperative to know where the data has been, and who has touched it. When was it changed, what are the chains of relationships, and how will the data be used elsewhere?

There is an obvious need to make AI predictions easier to trace and explain. Without this, the deployment of AI will likely slow down and prevent further adoption. This is especially important in healthcare and criminal justice cases. In these cases, the consequences of error are more significant than just delivering a poor customer experience. AI needs explainability to function effectively.

Introducing graph technology

Graph database technology can provide instant help. In a graph database, it is much easier to track how that data is changed, where data is used, and who used what data. It is a technology that has been used extensively for data lineage to meet data compliance regulations such as Europe’s GDPR or the California Consumer Privacy Act (CCPA).

In the context of AI applications, graph technology can tackle the explainable data issue using data lineage methods. Graph technology incorporates the context and connections that make AI more broadly applicable. Understanding and monitoring data lineage also guards against the manipulation of input data. This is common in areas such as corporate fraud. In cases of fraud, the significance of input data is common knowledge, so people can manipulate information to avoid detection.

This makes the entire system inherently untrustworthy. Graph database technology helps build that trust. Having explainable data means an organisation knows what data was used to train its model and why. To achieve this, it requires storing data as a graph database. The graph database will provide ‘explainable data’ on the three axes of data, predictions, and algorithms outlined earlier.

Tracking the ripple effects from data

Graph databases are also exceptional at tracking the chain of data change and any subsequent ripple effects. This brings a far deeper insight into data and its impact. Another area with potential is research into explainable predictions. In the case of explainable predictions, people want to understand what weights and features were used for a particular prediction. One example of this is the association of nodes in a neural with a labeled knowledge graph. When a neural network uses a node, the user can gain insight into all the node’s associated data from the knowledge graph.

Graph algorithms could one day also allow us to know which individual layers and thresholds can lead to a prediction. This is not something that will happen over the next few years. There is promising research in this area. Early examples include constructing a tensor in a graph database with weighted linear relationships. Promising early signs also indicate we may be able to determine explanations and coefficients at each layer, too.

High-profile mistakes in AI will erode public trust in it. This will focus brands’ attention on addressing the problem, given that for many, AI is integral to their future vision. The predictions and decisions made by AI need to be much more easily interpretable by experts and explainable to the public. If a foundation of trust can’t be built, people will instinctively reject recommendations that do not feel right. As a result of losing out on these recommendations, we miss out on ways to cut through information deluge and get to valuable or interesting information.

For AI to realise its full potential, graph technology will be essential. Without it, organisations will struggle to deliver the explainability factor required for AI to be truly trusted. And without that trusted AI, brands will pay the consequences.  

By Amy Hodler

Amy Hodler is Director, Analytics and AI Program at Neo4j, the world’s leading graph database company. She is also co-author of Graph Algorithms: Practical Examples in Apache Spark & Neo4j, published by O’Reilly Media.

Read more from Amy Hodler