Common Sense in Artificial Intelligence

In 2011, IBM’s Watson defeated human champions in a quiz game called jeopardy!. This success of AI creation against humans made Watson famous among AI enthusiasts. Since then the founder of IBM Watson, David Ferrucci has taken the AI and its advancement with more passion. Now, the man is up to something quite interesting and challenging and that is introducing common sense in Artificial Intelligence.

IBM Watson wins in jeopardy!

We think of Artificial Intelligence as a simulation of human intellect. We want machines to learn the human qualities and capabilities in all its finesse. This motivation has given rise to machines listening to our commands and resulted in self-driving cars. But all these AI advancements lack one thing. Common sense. Ironically, this ability of humans is the most frequently used one. We use it so often that it is unnoticeable. Yet this is missing from all AI systems till now. A search command may compel SIRI to scan through the internet to find an answer to your query but it is blank about what’s going on. The self-driving car may get cautious of avoiding obstacles on the road but it doesn’t get why saving a human life is so important.

common sense in Artificial Intelligence

David Ferrucci has taken upon this huge and complex task of making AI systems familiar with common sense. General knowledge about facts that humans ‘happen’ to know about. He has started with the process of storytelling. A story of two children, Fernando and Zoey, is narrated to the AI creation.

Fernando and Zoey bought some plants. Fernando keeps his plant in the window sill where it could get sunlight while Zoey puts it in the room. After a few days, Fernando’s plant is all grown up while Zoey’s is all withered. She then puts it in the window sill and it blooms.

The question then appears on the screen, “Does it make sense that Fernando put his plant in the window because he wants it to be healthy? The sunny window has a light and the plant needs to be healthy.”

Through questions, the AI system is trying to get hold of general knowledge that is obvious for us, humans. Like sunlight is necessary for a plant’s life.

When the person answers ‘yes’, this knowledge is stored as a fact by an AI program named CLARA. CLARA stands for Collaborative Learning and Reading Agent. CLARA then asks a series of questions related to plant story like an inquisitive learner. Through this question-answer process, the AI program is trying to grasp the basic knowledge that humans know about plants and sunlight!

To make full use of Artificial Intelligence, it is necessary that it reasons and applies logic like a human brain. It is the main requirement of what we require from AI. An NYU professor, Earnest David, explains this as, “central to most of what we want to do with AI. This is the large obstacle that the current approaches are having serious trouble with.

After AI win in Jeopardy, Ferrucci wanted to explore more the depths of Artificial Intelligence. After a year of AI win in Jeopardy!, he moved to Elemental Cognition which is funded by Bridgewater Associates. Ferrucci has been working on common sense in Artificial Intelligence for quite some time now.

Since jeopardy, Artificial Intelligence has evolved a lot too. Deep learning has given AI system a new edge. They are now able to detect faces and translate texts, by feeding large amounts of data to them. Mainly, deep learning has evolved much around language understanding. Feeding a neural network with large amounts of text enables it to learn text coherence and sentence formation. AI has produced amazing results in this aspect by successfully completing sentences with much soundness.

But when these AI models were given the plant story, they were unsuccessful in completing the sentence with some sound information. These advanced language models still lack the notion of common sense and reasoning. The sentence “Zoey moves her plant to a sunny window. Soon …” was given to be completed by the AI model. The result was a series of strange sentences like “she finds something, not pleasant,” “fertilizer is visible in the window,” and “another plant is missing from the bedroom.” This showed that the AI model tried completing the sentence through mere statistical patterns.

CLARA wants to go a step further that is combine statistical deep learning models and old fashioned logic rules. It uses statistical pattern matching to figure out placement of nouns and verbs and logic to figure out why an event happens and causes that further lead to new events.

As you may have guessed by now, that deep learning will make this complex process a lot easier. Instead of hand-engineering all basic knowledge of different aspects, large data comprising of general knowledge will be fed to the AI system. Learning this general knowledge will eventually lead Artificial Intelligence to produce its own general knowledge. Like from plant story, CLARA may learn that sunlight makes plant green.

According to MIT professor, Roger Levy, “It’s a very challenging enterprise, but I think it’s an important vision and goal. Language is not just a set of statistical associations and patterns—it also connects with meaning and reasoning, and our common sense understanding of the world.” Roger works in the domain of AI and cognitive science.

This whole process of introducing common sense in Artificial Intelligence is a long process despite the concept being attractive and interesting to research and work on. As David Ferrucci himself explains, “Can we ever get machines to actually understand what they read? That’s a very hard thing, and that’s ultimately what Elemental Cognition is about.

The assistance from common people and an investor with deep pockets may fasten the process. Scaling up a system with stories and basic knowledge is a difficult and slow process. Ferrucci suggests building a knowledge database by asking questions from students by giving them some text to read. In the end it all boils down to how much a common man is interested in helping a search engine or AI-powered common-sense knowledge database.

Elemental Cognition has not yet released any detailed research papers or source code related to this project. However, a research paper discussing machine comprehension and how it should be improved was published on 4th May 2020.

Some researchers like Yejin Choi are convinced that at a high level the research and working regarding common sense in Artificial Intelligence by Elemental Cognition do make sense. But still, she finds it difficult to believe an AI system with general common sense.

Some major research and breakthroughs are required before we can achieve this incredible quality of common sense in Artificial Intelligence models. Making machines understand and process information as seamlessly as humans is not an easy effort. As David from NYU says, “There seems to be something serious we’re missing. There are aspects of it that we haven’t gotten anywhere near“.

Leave a Reply

Your email address will not be published. Required fields are marked *