Let's consider this example:

It's John's birthday, let's buy him a kite.

We humans most likely would say the kite is a birthday gift, if asked why it's being bought; and we refer to this reasoning as common sense.

Why do we need this in artificially intelligent agents? I think it could cause a plethora of problems, since a lot of our human errors are caused by these vague assumptions.

Imagine an AI ignoring doing certain things because it assumes it has already been done by someone else (or another AI), using its common sense.

Wouldn't that bring human errors into AI systems?

  • $\begingroup$ Comments are not for extended discussion; this conversation has been moved to chat. $\endgroup$ – nbro Mar 7 '20 at 17:44
  • 1
    $\begingroup$ Why do human beings need common sense? $\endgroup$ – Basile Starynkevitch Feb 8 at 6:53

Commonsense knowledge is the collection of premises that everyone, in a certain context (hence common sense knowledge might be a function of the context), takes for granted. There would exist a lot of miscommunication between a human and an AI if the AI did not possess common sense knowledge. Therefore, commonsense knowledge is fundamental to human-AI interaction.

There are also premises that every human takes for granted independently of the country, culture, or, in general, context. For example, every human (almost since its birth) has a mechanism for reasoning about naive physics, such as space, time, and physical interactions. If an AI does not possess this knowledge, then it cannot perform the tasks that require this knowledge.

Any task that requires a machine to have common sense knowledge (of an average human) is believed to be AI-complete, that is, it requires human-level (or general) intelligence. See section D of AI-Complete, AI-Hard, or AI-Easy – Classification of Problems in AI (2012) by Roman V. Yampolskiy.

Of course, the problems that arise while humans communicate because of different assumptions or premises might also arise between humans and AIs (that possess commonsense knowledge).

  • $\begingroup$ Comments are not for extended discussion; this conversation has been moved to chat. $\endgroup$ – nbro Mar 7 '20 at 17:42

We need this kind of common sense knowledge if we want to get computers to understand human language. It's easy for a computer program to analyse the grammatical structure of the example you give, but in order to understand its meaning we need to know the possible contexts, which is what you refer to as "common sense" here.

This was emphasised a lot in Roger Schank et al.'s work on computer understanding of stories, and lead to a lot of research into knowledge representation, scripts, plans, goals. One example from Schank's work is Mary was hungry. She picked up a Michelin Guide. -- this seems like a non-sequitur: if you are hungry, why pick up a book? Until you realise that it is a restaurant guide, and that Mary is presumably planning to go to a restaurant to eat. If you know that going to a restaurant is a potential solution to the problem of being hungry, then you have no problem understanding this story fragment.

Any story needs common sense to be understood, because no story is completely explicit. Common things are "understood" and aren't explicitly mentioned. Stories relate to human experience, and a story that would make everything explicit would probably read like a computer program. You also need common sense to understand how characters in a story behave, and how they are affected by what is happening. Again, this is very subjective, but it is necessary. Some common sense might be generally applicable, other aspects of it won't be. It's a complex issue, which is why researchers have struggled with it for at least half a century of AI research.

Of course this would introduce "human errors" into an AI system. All this is very subjective and culture-specific. Going to a restaurant in the USA is different from going to one in France -- this is why going abroad can be a challenge. And my reading of a story will probably be different from yours. But if you want to simulate human intelligence, you cannot do that without potential human "errors".

  • $\begingroup$ Comments are not for extended discussion; this conversation has been moved to chat. $\endgroup$ – nbro Mar 7 '20 at 17:45

I'll answer this question in several parts:

Why do AGI systems need to have common sense?

Humans in the wild reason and communicate using common sense more than they do with strict logic, you can see this by noting that it is easier to appeal to someone's emotion than logic. So any system that seeks to replicate human cognition (as in AGI) should also replicate this tendency to use common sense.

More simply put, we'd wish that our AGI system can speak to us in common sense language simply because that is what we understand best (otherwise we wouldn't understand our friendly AGI would we?). Obtuse theory and strict logic might technically be correct, but don't appeal to our understanding.

Isn't the goal of AGI the create the most cognitively advance system? Why should the "most perfect" AGI system need to deal with such imperfections and impreciseness present in common sense?

First, it might only appear to be the case that common sense logic is "irrational". Perhaps there is a consistent mathematical way to model common sense such that all the subtleties of common sense are represented in a rigour fashion.

Second, the early study of Artificial Intelligence started in the study of cognitive science, where researchers tried to replicate "algorithms of the mind", or more precisely: decidable procedures which replicated human thought. To that extent then, the study of AI isn't to create the "most supreme cognitive agent" but to merely replicate human thought/behavior. Once we can replicate human behavior we can perhaps try to create something super-human by giving it more computational power, but that is not guaranteed.

I still don't see why common sense is needed in AGI systems. Isn't AGI about being the most intelligent and powerful computational system? Why should it care or conform towards the limits of human understanding, which requires common sense?

Perhaps then you have a bit of a misaligned understanding of what AGI entails. AGI doesn't mean unbounded computational power (physically impossible due to physical constraints on computation such as Bremermann's limit) or unbounded intelligence (perhaps physically impossible due to the prior constraint). It usually just means artificial "general intelligence", general meaning broad and common.

Considerations about unbounded agents are studied in more detail in fields such as theoretical computer science (type theory I believe), decision theory, and perhaps even set theory, where we are able to pose questions about agents with unbounded computational power. We might say that there are questions even an AGI system with unbounded power can't answer due to the Halting Problem, but only if the assumptions on those fields map onto the structure of the given AGI, which might not be true.

For a better understanding of what AGI might entail and its goals, I might recommend two books: Artificial Intelligence: The Very Idea by John Haugeland for a more pragmatic approach (as pragmatic as AI-philosophy can be, and On the Origin of Objects by Brian Cantwell Smith for a more philosophically inclined approach.

As a fun aside, the collection of Zen koan's: The Gateless Gate, includes the following passage: (quoted and edited from wikipedia)

A monk asked Zhaozhou, a Chinese Zen master, "Has a dog Buddha-nature or not?" Zhaozhou answered, "Wú"

Wú (無) translates to "none", "nonesuch", or "nothing", which can be interpreted as to avoid answering either yes or no. This enlightened individual doesn't seek to strictly answer every question, but just to respond in a way that makes sense. It doesn't really matter as to wether the dog has Buddha-nature or not (whatever Buddha-nature means), so the master defaults to absolve the question rather than resolving it.

  • 3
    $\begingroup$ It would have been much funnier if the master had answered "wúf wúf!" $\endgroup$ – m3characters Jul 9 '19 at 9:21

Perhaps it would help to give an example of what can go wrong without common sense: At the start of the novel "The Two Faces of Tomorrow" by James Hogan, a construction supervisor on the Moon files a request with an automated system, asking that a particular large piece of construction equipment be delivered to his site as soon as possible. The system replies that it will arrive in twenty minutes. Twenty minutes later, the supervisor is killed as the equipment crashes into his construction site. The system had determined that the fastest way to deliver the equipment to that site was to mount it on a mass-driver and launch it at the site. Had the system in question been given common sense, it would have inferred additional unstated constraints on the query, such as 'the equipment should arrive intact', 'the arrival of the equipment should not cause damage or loss of life', and so on. (the rest of the novel describes an experiment designed to produce a new system with common sense)

  • $\begingroup$ Quite interesting example. Though, maybe it could go the other way around as well. Imagine an AI checking in on a patient, and by seeing the patient's overall normal looks, conclude that the patient might/should be fine (based on it's common sense that if you look fine then probably you are); however, the patient is in need of serious help. However, I think it's clear by now that common sense is not the problem, it's faulty common sense/knowledge that might turn out to be dangerous in critical situations. So I guess we should control and make sure that the AI has a flawless common sense. $\endgroup$ – Titan Jul 10 '19 at 15:56

Is this common sense, or is this natural language understanding?

It's been said that natural language understanding is one of the hardest AI tasks. This is one of the examples showing why. The first part of the sentence is related to the second part, that how sentences work.

Now the relevant question is how the two parts are related. There are a few standard relations that we encounter, for instance a temporal order. In this specific example, the nature of the relation is closer to a cause-and-effect.

You see this effect when we insert a word to make this relation explicit:

It's John's birthday, so let's buy him a kite. or Let's buy John a kite, because it's his birthday.

This is a technique for humans to make these implicit relations explicit.

Now, as curiousdannii notes, you also need the cultural knowledge to understand how a birthdays can be a cause for a present. No amount of common sense helps with that.

  • $\begingroup$ We use the "common sense" we have to understand a sentence like that. NLP is really hard, I agree, and that's because so far AI hasn't possessed this comprehensive knowledge and model of the world that we humans have; hence it lacks in connecting the dots and understanding those implicit relations. That "cultural knowledge" you mentioned is basically considered our common sense. $\endgroup$ – Titan Jul 9 '19 at 15:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.