In our lives, we meet different people and describe their common sense based on how they act on a situation. For example, highly extrovert people are able to deal with people without any awkwardness. For them, an action in how to deal with people come as common sense. But, in the case of scientists, approach to solving a problem may be common sense which ordinary people cannot see.
How can we define common sense in an AI agent?
I define "common sense" inline with human beings, to concatenate to intelligent agents;
An algorithmic agent, which has the ability to solve effective decision in relation to the way humans perceive their environment and common situations.
According to McCarthy John,
The first artificial intelligence program proposed to address common sense was Advice Taker
Currently, common sense is unsolved problem in AI. For more information, see John McCarthy's Programs with Common Sense.
I came up with a few ideas I would argue are valuable in motivating the idea of common sense for a machine learning model.
Common sense is retrospective. We define it in terms of a past (sensible) actions and conditions, and we can say someone has good common sense on the basis of their behavior, which we can view as the sum of their historical actions and the degree to which they were sensible.
The actions alone are not sufficient for demonstrating common sense; the mental model(s) that generated those actions is important as well. Why did the individual take those sensible actions? Is their reasoning behind their action sensible as well, or did they merely get lucky (put another way, does their generally not sensical mental model produce sensible actions in certain cases and non sensical actions in most others)? This hints that common sense is contingent on the rationalization of the actions themselves.
Given the former, common sense is contingent on the enumeration of possible actions and choosing the right one, given the context of the situation. For example, I’m walking on a path and see a snail. What are some actions? I could keep walking, stop and admire it for a while, step on it, or eat it. The first two options are sensible if I value the snail as a living being. The first option is sensible if I’m in a rush. The second and fourth may be sensible if I’m a chef exploring nature for new potential ingredients. The third option if I recognize the snail as invasive. We say someone has common sense when, given the context, the actions chosen are sensible and we can reason about them intuitively.
My guess is that, the intuition behind the perceived sense of an action is what you’re after. I’d argue that, ultimately, the intuition of common sense is defined by the person developing the model and will have to do with the formulation of the model (e.g it’s assumptions, the objective function, etc). After all, common sense is subjective and context specific.
Concretely, a model can have common sense if the developer bakes it in and we can use inferential methods to demonstrate this. For example, in a Word2Vec model, we might see that $\mathsf{Paris} \mapsto \mathsf{France}$ and would expect that $\mathsf{Tokyo} \mapsto \mathsf{Japan}$. To interrogate this, we might do some vector math and find that $\mathsf{Paris} - \mathsf{France} + \mathsf{Tokyo} = \mathsf{Japan}$. How and if the AI model develops the larger association between $\mathsf{Capital} \mapsto \mathsf{Country}$, however, comes down to how the developer built and trained the model to recapitulate their own common sense.
My sense is that common sense tends to be axiomatic. To avoid pitfalls, a degree of wisdom may also be required in that axioms may not apply in all contexts. [See Axiomatic System].
A major problem is that science often demonstrates that intuition, and "common sense", often lead to incorrect conclusions. Neil Degrasse Tyson covers this topic for the general public in his book Death By Black Hole:
Chapter 3, "Seeing Isn't Believing", hints at the pitfalls of generalizing from too little evidence. It begins by making the point that although we know the Earth is round, it appears flat when one observes only a small, local portion of it.
A very famous example comes from mathematician Abraham Wald:
During World War II, Wald was a member of the Statistical Research Group (SRG) where he applied his statistical skills to various wartime problems. These included methods of sequential analysis and sampling inspection. One of the problems that the SRG worked on was to examine the distribution of damage to aircraft to provide advice on how to minimize bomber losses to enemy fire. There was an inclination within the military to consider providing greater protection to parts that received more damage but Wald made the assumption that damage must be more uniformly distributed and that the aircraft that did return or show up in the samples were hit in the less vulnerable parts. Wald noted that the study only considered the aircraft that had survived their missions—the bombers that had been shot down were not present for the damage assessment. The holes in the returning aircraft, then, represented areas where a bomber could take damage and still return home safely. Wald proposed that the Navy instead reinforce the areas where the returning aircraft were unscathed, since those were the areas that, if hit, would cause the plane to be lost.
Source: Abraham Wald (wiki)
My sense is that "confidence levels" may be the main technique driving toward algorithmic "common sense", specifically in that the algorithm is questioning it's assumptions.
