We live in the age of artificial intelligence — and, in addition to your plastered conversations with Alexa in the living room, there is still a lot to be built in this field. One of the main issues, according to a researcher and assistant professor at the University of Southern California, is to make machines develop common sense. “It’s extremely challenging,” says Mayank Kejriwal.
Think about it — software development is full of codes and sets of equations that result in responses and stimuli. It is possible to program a machine to respond whenever you ask what time it is, or who founded Microsoft. These are predictable questions that can be quickly addressed to the databases that feed the program in question.
As machine learning evolves, more complex equations are put to the test, in the most specific situations you can imagine. However, this is exactly where the problem lies: there is no specificity for common sense.
“Modern AI is designed to deal with highly specific problems, in contrast to common sense, which is vague and cannot be defined by a set of rules. Even the latest models make absurd mistakes at times, suggesting that something fundamental is missing from the AI world model.”
Is it possible to calculate common sense?
The professor explains that true common sense, as developed by humans, is very difficult to calculate. It’s a natural skill, not formally taught—and even for us it can be a challenge (see how many humans without “standard common sense” are out there).
Based on a series of conditions inherent to our existence, this ability is linked to life in society, but not only it – there are unconscious learning of physics and abstract knowledge, such as time and space, which are only acquired… Living.
For all that, Kejriwal explains that one of the most important challenges of artificial intelligence is to break this boundary between the apparent common sense of machines and human common sense. Of course, right now there are several scientists and institutions working to understand how to pass this ability on to robots. The researcher cites the Machine Common Sense program, launched by the US Defense Advanced Research Projects Agency, as an example.
“The Machine Common Sense program funds many current research efforts in machine common sense, including our own, Multi-modal Open World Grounded Learning and Inference (MOWGLI). MOWGLI is a collaboration between our research group at the University of Southern California and AI researchers from the Massachusetts Institute of Technology, the University of California at Irvine, Stanford, and the Rensselaer Polytechnic Institute. The project aims to build a computer system that can answer a wide range of common sense questions.”
But if common sense to some humans seems like a lost cause, Kejriwal says there are good prospects for machines. With the advanced deep learning AI dubbed “Transformers” (or “Transformers” in Portuguese), it is possible to model natural language in a highly efficient way, which allows to answer some simple questions about common sense. This was one of the big steps to allow robots to communicate in a more human way.
The evolution of transformers has led the scientific community to question deeper and even philosophical topics. After all, who can establish what common sense is? And how do you know if machines are really acquiring the skill?
“Researchers divide common sense into different categories, including common sense sociology, psychology and prior knowledge. Authors of a recent book argue that researchers can go much further, dividing these categories into 48 fine-grained areas such as planning, threat detection, and emotions. However, it is not always clear how these areas can be neatly separated. Experiments have suggested that a clear answer to the first question can be problematic. Even expert human annotators—people who analyze text and categorize its components—within our group disagreed about which aspects of common sense applied to a particular sentence. Annotators agreed on relatively concrete categories such as time and space, but disagreed on more abstract concepts.”
In conclusion, Kejriwal states that despite all the evolution, this area of artificial intelligence is still extremely uncertain and needs many further studies. For him, common sense is linked to so many variants and learning paths that it is impractical for a machine (or the AI field, so to speak) to reach a comfortable result for this impasse without walking all this same road.
“Depending on the success of the new lines of research [de aprendizado profundo, focadas em senso comum], there is no way to say what the common sense of the machines will be in five years, or in 50”, completes the researcher.
With information: The Conversation