Op-ed: AI needs to understand common sense

Mohamed Suliman, contributor

People usually don’t ponder over many decisions they make. We’re capable of judging if a piece of furniture will fit through a doorway or discerning people’s intentions from the look on their faces. We can tell if they agree with us or are uncomfortable with something we said. Common sense is the cognitive capacity to perform all these decisions, and it’s crucial to navigating life. Researchers of artificial intelligence, or AI, are looking for ways to equip machines with the capacity of common sense by understanding and formalizing how humans use common sense. 

There is no unified definition of common sense or its source; the philosopher Aristotle defined it as the ability to discriminate against particular objects. The Collins dictionary defines it as the “natural ability to make good judgments and to behave in a practical and sensible way.” There are also scientific attempts to define common sense with concepts such as naive physics, which explains the ability to understand and judge physical objects around us and, more broadly, understand the concept of time and space. The concept of common sense also includes the theory of mind, which explains how humans can infer the intentions of others and predict their actions. 

There is an ongoing debate among scientists about the origin of common sense in humans. Some cognitive scientists believe humans are born with core knowledge and have largely innate mental abilities. Other researchers argue some of these skills are found early, even in infants and babies, but are refined through development. Part of common sense could be built in while the other part is not. The origin of the common sense debate is part of the long-standing nature versus nurture debate. 

Most current AI systems rely on classification and pattern recognition models that attempt to capture and abstract the notion of intelligence; albeit their known success in many fields, they still need an understanding of common sense. For example, a 2020 study showed the natural language processing systems couldn’t reason enough to understand the deeper meaning behind some statements that require an understanding of the common sense behind them. Researchers also revealed how machines still can’t use abductive reasoning, which is the ability to reason when faced with a scene of incomplete information. This is a skill humans use on a daily basis to solve problems.

This research goes deep first to understand the essential nature of common sense to formalize it. One of the research directions is understanding spatial and physical reasoning. Some researchers started thinking about a machine’s common sense in a physical setting. This research aims to help the machine reason about the whole scene and the semantic relationship between the objects. Other directions aim to build a knowledge database about the world that could be used by the machine, especially in problems with incomplete information. 

To implement AI common sense practically, different frameworks and programming languages are proposed and used. Crowdsourcing common sense knowledge projects such as Wordnet, Cyc and Concept Net, has shown to be successful in many aspects. Mining the web is also another way to extract common sense automatically. Most recently, researchers started creating probabilistic programming languages such as Gen and Church to encode common sense into a language the machine understands.

There have been several attempts to incorporate common sense into AI. In 1959, the advice-taker system was introduced as a hypothetical system. It paved the way for the rest of the subsequent attempts to incorporate common sense into AI. Today, these systems have matured to answer questions about issues related to day-to-day life and reason about visual events. The Defense Advanced Research Projects Agency, or DARPA, also initiated a project called machine common sense with two strategies, one learning from experiences and the other learning from scanning the web to create common sense knowledge. The Multi-Modal Open World Grounded Learning and Inference, or MOWGLI, is a cross-university collaboration project to build a system that can answer a wide range of common sense questions. 

Companies and startups also started considering the common sense approach. Element Cognition, an AI startup founded in 2015, uses a dataset to let AI reason about the common sense behind stories. DeepMind also built a system that mimics how infants reason about objects around them. 

Many tests are now proposed to measure the ability of the AI system’s common sense capacity. For instance, IBM developed AGENT, a benchmark that measures AI agents’ abilities to exhibit core psychological principles such as inferring other intentions and differentiating objects from humans. However, some researchers argue  we need a more robust understanding of common sense’s complexity, such as probabilistic judgment of events, comprehension of information such as stories and enumeration and evaluation. These tests are of crucial importance to discern real progress from the hype in the field. 

Building AI with common sense capabilities is guided by the philosophy that the ideal AI system should behave and reason like a human. The field of common sense AI is still in its early days. To achieve significant success and build real-world applications, investors should take more risks and invest in this growing sector. While different from the existing AI models, researchers from other AI fields should understand that this approach complements pre-existing approaches to AI rather than competing with them. The implementation of common sense capabilities in AI is instrumental in the development of making technology more human.

Mohamed Suliman is a senior researcher at Northeastern University’s Civic AI Lab, he also holds a degree in Mechanical Engineering from the University of Khartoum, and he tweets at @MuhammedKambal