Created by Materia for OpenMind Recommended by Materia
6
Start Common Sense Knowledge, Crucial for the Success of AI Systems
25 February 2020

Common Sense Knowledge, Crucial for the Success of AI Systems

Estimated reading time Time 6 to read

Common sense is the knowledge that all humans have. Such knowledge is unspoken and unwritten – we take it for granted. We acquire it imperceptibly from the day we are born. For example, “animals don’t drive cars” or “my mother is older than me”. This knowledge is often used by human experts even when solving very narrow, domain-specific tasks. This common-sense knowledge is something that we learn through experience and curiosity without even being aware of it. We also acquire a great deal of it in our lifetimes.

BBVA-OpenMind-ESCHER-Common sense knowledge-IA
M.C. Escher, Relativityy  Source: Sharon Mollerus. Flickr

AI systems do not have common sense knowledge and acquiring it has been seen as important since their beginning. Furthermore, from all the efforts made over many years, it’s become evident that building common sense reasoning systems is a work-intensive and sometimes costly task. In this paper, I show why common sense reasoning is so important, and describe some approaches that have been used to build these systems. These approaches have enabled common sense reasoning tasks to be used as add-ons to AI client programs – such as Chatbots.

Why is common sense knowledge so important for AI systems?

One of the founding fathers of AI, John McCarthy, was amongst the first to realize its importance. He wrote a paper that was the first to propose common sense reasoning through a hypothetical program called Advice Taker in 1959. This paper only described a specification for what a common-sense program should do. However, it soon became apparent that there was a need for working common sense knowledge programs to assist decision making in AI expert systems. These systems represented the first commercial boom period for AI, so common sense knowledge became seen as an essential adjunct for their success.

The reason why common sense reasoning was seen as important was that many of these systems were very competent at problem-solving, but they were also brittle, because, they often gave meaningless answers when trying to reason with unusual problem data. For instance, as I show later, a medical diagnostic expert system didn’t realize that it was given car data when it diagnosed a car with measles. These absurdities would never happen using human experts because their common sense would enable them to know it was invalid data. Also, human experts sometimes reason in ways in which common sense might even override their specialist domain knowledge. For example, a medical expert might be considering life-threatening surgery on a very elderly patient with a short life expectancy. The expert may rule out the surgery given his or her common sense understanding of the circumstances. An AI expert system, without common sense knowledge, would not have this broader perspective to make such a decision.

Why are some deep-learning projects struggling to adapt to change?

In my last OpenMind publication, I discussed the difficulties with delivering human-level AI using deep learning. Whilst deep learning excels at pattern recognition, it’s very poor at adaptation even when small modifications are encountered, and has to be re-trained with large amounts of data from scratch. In contrast, humans can adapt with more ease by engaging in reasoning and making logical inferences from common sense knowledge. This means that when a human expert encounters changes in a domain or new domains that have some familiarity, they can often adapt in ways in which deep reasoning models cannot.

BBVA-OpenMind-Keith Darlingon-Human level Artificial General Intelligence AGI
Human-level intelligence has come to be known as Strong AI or Artificial General Intelligence (AGI).

Some common sense reasoning projects

A number of common sense projects have been developed over the last 30years. The first was CYC [1] that started in 1984. This was a very ambitious AI project that attempted to represent common-sense knowledge explicitly by assembling an ontology of familiar common sense concepts. It was named CYC taken from the word ENCYCLOPAEDIA because it was attempting to store a vast amount of real-world knowledge. The project was originally based at the Microelectronics and Computer Technology Corporation (MCC) in Austin, Texas. The project leader and founder Doug Lenat was quoted as saying: “The motive behind CYC is that the first generation of expert systems lack common sense. This has caused them to be “brittle”. This has meant that many first-generation expert systems performed very well but only when confined to the narrow boundaries of the domain knowledge used by the systems. The first-generation expert systems would be unable to respond sensibly to broader real-world common sense knowledge which human experts frequently use because they didn’t have any”. As an example of this problem, Lenat considers the following skin disease diagnosis expert system dialogue with a patient who is actually a Chevrolet car.

Program: Are there any spots on the body?

User: Yes.

Program: What colour spots?

User: Reddish brown

Program: Are there more spots on the trunk than elsewhere?

User: No.

Program: The patient has measles.


Common sense will help machines not to fall into obvious errors for humans

This example shows that AI systems can produce bizarre output when confronted with unexpected situations. Lenat calls this the breadth hypothesis. This states: “To behave intelligently in unexpected situations, an agent must be capable of falling back on increasingly general knowledge and analogising to specific but far-flung knowledge”. Many years of effort went into CYC development by creating logic relationships (sometimes called predicates) with millions of facts about the world. The purpose of CYC was to enable AI applications to use their domain-specific knowledge and perform human-like common sense reasoning. It has been operational for many years and has many commercial customers.

According to the CYC Website “CYC has the world’s broadest and deepest common sense knowledge base (KB), by orders of magnitude. The knowledge base is not a database – it consists of real-world axioms that CYC uses to reason about the world and understand your data. CYC’s KB includes more than 10,000 predicates, millions of collections and concepts, and more than 25 million assertions”. However, there were shortcomings identified with the CYC project – particularly dealing with the ambiguities of human language. It also has its share of critics. According to [2], they think the goals of the project are admirable, but CYC is still too incomplete on its own to have an enormous impact. The project required a huge amount of handcrafted labour – meaning there are many expert staff required to develop this system and was estimated to have cost more than $200 million to develop.

Other less costly approaches have used an open-source model for capturing data on the Web. Open-source means that a community of users who come to the Website can help in its construction. For example, ConceptNet is a freely available common sense knowledge-base and natural language processing toolkit – designed to help computers understand the meanings of words that people use. It does this by using a means of representing knowledge called, semantic networks. These use graphical methods to describe relationships between concepts and events to describe common sense activities. This semantic network knowledge-base contains over 1.6 million assertions of common sense knowledge such as the spatial, physical, social, and psychological aspects of everyday life. It was launched in 1999 at the MIT Media Lab and began as a system called Open Mind Common Sense. It has been successfully used in Chatbots and some natural language support.

Another approach has been to get the computer to learn to read on the Web (called Web scraping) to acquire common sense knowledge. One such effort has been through a system developed at Carnegie Mellon University that began in 2011 called NELL (Never Ending Language Learner). The basic idea here is that NELL is programmed to search the Web and identify linguistic patterns to deduce its meaning. It reports on its recently-learned facts via its Website. For example, it recently learned that Oliver Stone contributed to the film JFK, and that day lily leaves are a vegetable. These are, perhaps, not particularly exciting finds but nevertheless, contribute to a knowledge-base of over 50 million beliefs with, they say, high confidence.

DARPA, the US defense department’s research agency, has also recognized the absence of common sense as being an important issue. They recently launched a project called Machine Common Sense. As they say:“ The absence of common sense prevents intelligent systems from understanding their world, behaving reasonably in unforeseen situations, communicating naturally with people, and learning from new experiences. Its absence is considered the most significant barrier between the narrowly focused AI applications of today and the more general, human-like AI systems hoped for in the future”. DARPA’s approach is to use a two-pronged strategy. The first strategy in their proposal involves building computational models that learn from experience. The second strategy seeks to develop a service that learns from reading the Web, like a research librarian, to construct a common sense knowledge repository.

Conclusions

Despite many valiant efforts, there is a general feeling that insufficient progress has been made in common sense applications for AI. One of the problems is that it is very difficult to formulise because it is a very messy unstructured domain. It is also difficult to know what constitutes completeness because of the lack of a precise definition of exactly what common sense is. However, progress is crucial if we are to overcome the problems described in this paper.

References

[1] Lenat, D. and Guha, R. Building Large Knowledge-Based Systems: Representation and Inference in the Cyc Project. Addison-Wesley 1990.

[2] Marcus, G. and Davis, E. Rebooting AI. Building Artificial Intelligence we can trust. Pantheon Books, 2019.

Keith Darlington

Comments on this publication

Name cannot be empty
Write a comment here…* (500 words maximum)
This field cannot be empty, Please enter your comment.
*Your comment will be reviewed before being published
Captcha must be solved