Ronald J. Brachman and Hector J. Levesque, Machines Like Us: Toward AI with Common Sense
Henry Kautz · Prometheus · 2023
Since its birth in the 1940s, the field of artificial intelligence has been divided into two camps, one focused on artificial neural networks and the other on reasoning with symbolic representations of knowledge.The symbolic representational approach firmly dominated the field until 2012, when a neural network named 'AlexNet' handily won an algorithm competition for recognizing objects in images (Krizhevsky et al., 2012).Further convincing successes of neural network algorithms for speech recognition, the game of Go (Silver et al., 2017), and other problems that had long eluded the KR approach soon followed.Today, neural networks, under the banner of 'deep learning', where 'deep' refers to the fact that the artificial neurons are arranged in many layers, dominate research and commercial applications.Most students studying AI learn little about knowledge representation, and the approach is rarely mentioned in news stories and popular accounts of AI.It would be a grave error, however, to conclude that the more than 50 years of research in knowledge representation and reasoning yielded no insights about the nature of computational intelligence.We are beginning to see deep learning researchers struggling with issues that have long been studied in knowledge representation, such as the nature of the concepts and categories that an AI system must employ to make sense of the world -an issue sometimes identified as 'learning disentangled representations' (Bengio et al., 2013).Furthermore, even today's most sophisticated deep learning systems are unreliable, in that they can unexpectedly and catastrophically fail on certain inputs.For example, at the time of writing this review, OpenAI's ChatGPT is the most powerful natural language processing system ever created, and its fluency in producing text has led some people wrongly to conclude that it must be conscious.It is not difficult, however, to find examples when ChatGPT produces plausiblesounding but entirely erroneous 'hallucinated' answers (Hofstadter, 2022).One path to trustworthy AI may be to base systems on explicit and validated representations of what they know and do not know (Marcus and Davis 2019).It is therefore timely that Machines Like Us: Toward AI with Common Sense has appeared to remind us of the insights developed by the knowledge representation research community and to restate fundamental open problems that cannot even be stated without reference to how an intelligent being represents and reasons about the world.Ronald Brachman and Hector Levesque are two pre-eminent researchers in the knowledge representation and reasoning tradition, individually and jointly publishing many influential papers from the 1980s through the present day.Brachman is best known for showing how so-called 'semantic networks', a graph-based knowledge representation approach that had been developed in the earliest days of AI and which is still in widespread use (including by Google and Facebook under the name of 'knowledge graphs') could be viewed as object-oriented versions of first-order logic (Brachman and Schmolze, 1985).Levesque collaborated on some of this work and went on to develop logic-based methods for reasoning about beliefs and plans (Levesque et al., 1997).In 2012, Levesque and two other collaborators, Ernie Davis and Leora Morgenstern, introduced Winograd schemas, a way to create challenge problems for natural language processing systems that could only be disambiguated by using commonsense knowledge; these problems have become a part of standard benchmark sets for such systems (Levesque et al., 2012).The book begins by defining what the authors mean by commonsense: 'the ability to make effective use of ordinary, everyday, experiential knowledge in achieving ordinary, everyday, practical goals'.They note that commonsense is related to but is more specialized than rationality, which