In this blog series, we discuss why Knowledge Graphs are excellent at guiding and focusing ML and at serving as a unifying fabric for different AI algorithms.
Artificial Intelligence (AI) and Machine Learning (ML) and Cognitive Computing are umbrella terms for a wide set of algorithms, technologies and approaches that make software seem ‘smart’. Such algorithms can discern patterns in data so that when new data comes in, they can apply patterns to make conclusions about new data.
• Knowledge representation and reasoning (KR) is the eld of AI dedicated to represent- ing information about the world in a form that a computer system can utilize.
• Knowledge Graphs are part of the KR branch of AI and can capture data as well as semantics or the meaning of data. They enable computers to reason based on the full available contextual and conceptual information.
• Leading companies who are building Knowledge Graphs include Google, Apple, Amazon, Airbnb, Bloomberg, Facebook, LinkedIn, Thomson Reuters — and these are just a few.
• As an enterprise Knowledge Graph infrastructure, TopBraid EDG supports Data Governance 2.0 and applications of AI/ML.
• Within TopBraid EDG, we show several examples of the application of AI to help automate the governance of information.
• We conclude with some thoughts on what governance is needed for AI itself to support its ongoing use in the enterprise.
The topic of Artificial Intelligence (AI) has become dominant as a technology trend and is influencing almost every industry. Today, you are probably hearing this term daily as it is often discussed in the news and featured in ads promoting different technology-enabled products.
Applications using AI can be found in all industries and they span different functions. Use of AI promises to deliver results that could not be achieved with more traditional technologies or to achieve the results cheaper and faster. At the same time, many organizations are just beginning their AI journey.
What is AI, what can it do, and what is hype vs. reality are questions that are not yet well understood.
What Is Artificial Intelligence?
Artificial Intelligence is an umbrella term for a wide set of algorithms, technologies and approaches that make software seem ‘smart’— giving us an impression of human-like smartness. Other terms for AI include Machine Intelligence and Computational Intelligence.
Artificial Intelligence can be divided into broad (or general) and narrow (or weak) AI.
Broad AI, where a machine could success- fully perform any intellectual task, does not exist today. Leading scientists agree that it is at least 30 years away. And many doubt if it will ever be possible.
What we see and use today are many examples of narrow AI that can carry out specific tasks without being explicitly programmed in a traditional way. Such examples include facial recognition in the security systems, purchase recommendations on e-commerce sites and cars that park themselves or inter- act with drivers using natural language.
A lot of specialized work goes into creating software capable of accomplishing one of these focused tasks. ML, reasoning, computer vision, knowledge representation, natural language processing, robotics — are all technologies and approaches that are part of AI.
AI algorithms are integrated into variety of applications and systems. What we consider to be AI changes over time. Once new software becomes widely available and routinely used, people typically stop thinking of it as AI.
What Is Machine Learning?
Machine Learning (ML) is a subset of AI that itself is also an umbrella term for multiple technologies. One common theme across all ML algorithms is that they discern patterns in data so that when new data comes in, they can apply patterns they know (have learned) to make conclusions about new data.
ML technologies include decision trees, Bayesian belief networks, k-nearest neighbors, case-based reasoning, neural networks and other approaches.
ML uses training data and the quality of its conclusions is highly dependent on the training data. Supervised ML learning relies on having paired input subjects and desired outputs in the training datasets. These approaches work well for classification/categorization tasks. Also, supervised ML is more likely to make judgments that humans can relate to — because humans have provided the basis for decisions. A system trained on specific data, for a specific task, will not perform well if data changes.
Unsupervised ML is trained on information that does not identify outcomes. The algorithms, often statistical, act on the information without guidance. It may, for example, nd “outliers” or unusual data. Results are less predictable, but can work with more general data.
What Is Knowledge Representation?
Knowledge representation and reasoning (KR, KR2, KR&R) is the field of AI dedicated to representing information about the world in a form that a computer system can utilize to solve complex tasks.
In the field of Knowledge Representation one typically distinguishes between:
- Knowledge Bases containing facts about the world and rules about the facts i.e., the knowledge; and
- Inference engines that perform reasoning by applying the rules to the available information in order to answer questions and solve problems.
Systems that use KR are sometimes called knowledge-based or expert systems.
Knowledge Representation requires a language in which we can say things about the world. There is a vast amount of knowledge available for capture. No single person or group can represent all of it —
just like no single entity can create all the information on the web. However, collectively, we can make a lot of knowledge available in a machine processable way — as long as the KR language is standardized and lets each of us work on our field of knowledge, but connect it with facts and rules developed by others — just like we do on the web.
KR separates knowledge from the conventional procedural code making it easier to define and maintain complex software logic and letting all stakeholders more readily share business rules.
What Is Cognitive Computing?
Generally speaking, Cognitive Computing systems try to simulate human thought processes. Cognitive Computing is often described as simply marketing jargon. Still, some working definitions and understanding of the differences between Cognitive Computing and AI have emerged.
Cognitive Computing combines multiple AI techniques and algorithms (such as ML and KR) in a way that makes it possible for machines to reason and understand complex situations.
In other words, Cognitive Computing uses many of the same fundamentals as AI to mimic the problem solving processes used by humans. AI, on the other hand, does not try to mimic human thought processes. Instead, a good AI system simply uses the best possible algorithm for solving a given problem.
Another point of difference is that Cognitive Computing is positioned as augmenting human capabilities and providing advise to humans — supplementing our own decision-making. While AI is positioned as a way to automate processes.
What Are Knowledge Graphs?
Knowledge Graphs represent knowledge, thus, are part of the KR branch of AI. They enable computers to reason based on the full available contextual and conceptual information. Knowl- edge Graphs are also excellent at guiding and focusing ML and at serving as a unifying fabric for the different AI algorithms.
A knowledge graph is self-descriptive, or, simply put, it captures the data as well as its meaning or semantics. As a result, computers can reason — derive additional implicit information from the explicitly stated information.
Different AI algorithms can use the knowledge graph — both, as input and as output. Results of processing come together, enriching each other. For example, given a training set with the information about people and provided with directions to learn about specific properties such as eye color or parents,
ML algorithms can learn that each person has no more than two parents, and that children of blue eyed parents always have blue eyes. These rules can then be captured explicitly in the knowledge graph as a suggested rule and checked by the human curators.
Once approved for inclusion, rules could be used to identify issues with missing or incorrect data. For example, to flag as an issue a blue eyed person whose parents have brown eyes. Established rules can also be used as “known axioms” in further application of ML to ensure that insufficient or incorrect training data does not cause erroneous conclusions.
Join us for Part 2 of our blog series to learn how Knowledge Graphs form a trusted foundation for AI and ML and which industry leaders have invested in Knowledge Graphs.