Artificial intelligence is expected to create trillions of dollars of value across the economy. But as the technology becomes a part of our daily lives, many people are still skeptical. Their main concern is that many AI solutions work like black boxes and seem to magically generate insights without explanation.
At the same time, knowledge graphs have been recognized by many industries as an efficient approach to data governance, metadata management and data enrichment and are increasingly being used as data integration technology. But knowledge graphs are also more and more identified as the building blocks of an AI strategy that enables explainable AI through the design principle called human-in-the-loop (HITL).
Why does artificial intelligence often work like a black box?
The promise of the AI, which is based on algorithms of machine learning such as deep learning, is to automatically extract patterns and rules from large datasets. This works very well for specific problems and, in many cases, helps automate classification tasks. Why exactly things are classified in one way or another cannot be explained. Because machine learning cannot extract causalities, it cannot reflect on why certain rules are extracted.
Machine learning algorithms learn from historical data, but they cannot derive new insights from it. In an increasingly dynamic environment, this is causing skepticism because the whole approach of deep learning is based on the assumption that there will always be enough data to learn from. In many industries, such as finance and healthcare, it is becoming increasingly important to implement AI systems that make their decisions explainable and transparent, incorporating new conditions and regulatory frameworks quickly. See, for example, the EU’s guidelines on ethics in artificial intelligence.
Can we build AI applications that can be trusted?
There is no trust without explainability. Explainability means that there are other trustworthy agents in the system who can understand and explain decisions made by the AI agent. Eventually, this will be regulated by authorities, but for the time being, there is no other option than making decisions made by AI more transparent. Unfortunately, it’s in the nature of some of the most popular machine learning algorithms that the basis of their calculated rules cannot be explained; they are just “a matter of fact.”
The only way out of this dilemma is a fundamental reengineering of the underlying architecture involved, which includes knowledge graphs as a prerequisite to calculate not only rules, but also corresponding explanations.
What is semantic AI, and what makes it different?
Semantic AI fuses symbolic and statistical AI. It combines methods from machine learning, knowledge modeling, natural language processing, text mining and the semantic web. It combines the advantages of both AI strategies, mainly semantic reasoning and neural networks. In short, semantic AI is not an alternative, but an extension of what is currently mainly used to build AI-based systems. This brings not only strategic options, but also an immediate advantage: faster learning from less training data, for example to overcome the so-called cold-start problem when developing chatbots.
What is a knowledge scientist?
Semantic AI introduces a fundamentally different methodology and, thus, additional stakeholders with complementary skills. While traditional machine learning is mainly done by data scientists, knowledge scientists are the ones who are involved in semantic AI or explainable AI. What is the difference?
At the core of the problem, data scientists spend more than half of their time collecting and processing uncontrolled digital data before it can be explored for useful nuggets. Many of these efforts focus on building flat files with unrelated data. Once the features are generated, they begin to lose their relationship to the real world.
An alternative approach is to develop tools for analysts to directly access an enterprise knowledge graph to extract a subset of data that can be quickly transformed into structures for analysis. The results of the analyses themselves can then be reused to enrich the knowledge graph.
The semantic AI approach thus creates a continuous cycle of which both machine learning and knowledge scientists are an integral part. Knowledge graphs serve as an interface in between, providing high-quality linked and normalized data.
Does this new AI approach lead to better results?
Apart from its potential to generate trustworthy and broadly accepted explainable AI based on knowledge graphs, the use of knowledge graphs together with semantically enriched and linked data to train machine learning algorithms has many other advantages.
This approach leads to results with sufficient accuracy even with sparse training data, which is especially helpful in the cold-start phase, when the algorithm cannot yet draw inferences from the data because it has not yet gathered enough information (see also: zero-shot learning). It also leads to better reusability of training datasets, which helps to save costs during data preparation. In addition, it complements existing training data with background knowledge that can quickly lead to richer training data through automated reasoning and can also help avoid the extraction of fundamentally wrong rules in a particular domain.
Developing An Interest In Semantic AI
If you are a data scientist or data manager — or if you manage someone in such a position — it’s important to start digging into this research and developing the skills to work with semantic AI.
Semantically enriched data serves as a basis for better data quality and offers more opportunities for feature extraction. This leads to higher accuracy in prediction and classification, calculated by machine learning algorithms. Furthermore, semantic AI should create an infrastructure to overcome information asymmetries between AI system developers and other stakeholders, including consumers and policymakers. Semantic AI ultimately leads to AI governance that works on three levels: technical, ethical and legal.
Most ML algorithms work well with either text or structured data. Semantic data models can close this gap. Relationships between business and data objects can be made available for further analysis. This allows you to provide data objects as training datasets composed of information from structured data and text.
Credit: Google News