Preface
- PDF / 171,413 Bytes
- 5 Pages / 439.642 x 666.49 pts Page_size
- 72 Downloads / 210 Views
Preface ¨ ´ 4,5 Andrew S. Gordon1 · Rob Miller2 · Leora Morgenstern3 · Gyorgy Turan
© Springer Nature Switzerland AG 2020
A few years after the 1956 Dartmouth Summer Workshop [1, 2], which first established artificial intelligence as a field of research, John McCarthy [3] discussed the importance of explicitly representing and reasoning with commonsense knowledge to the enterprise of creating artificially intelligent robots and agents. McCarthy proposed that commonsense knowledge was best represented using formal logic, which he viewed as a uniquely powerful lingua franca that could be used to express and reason with virtually any sort of information that humans might reason with when problem solving, a stance he further developed and propounded in [4, 5]. This approach, the formalist or logic-based approach to commonsense reasoning, was practiced by an increasing set of adherents over the next several decades [6, 7], and continues to be represented by the Commonsense Symposium Series, first held in 1991 [8] and held biennially, for the most part, after that. The commonsense reasoning landscape has changed considerably over the years. More than thirty years ago, Drew McDermott [9] noted that correctly specifying commonsense knowledge within formal logic was an error-prone enterprise, pointing to the brittleness of existing formal theories of common sense and the difficulty of modifying such theories in the face of new information. Statistical approaches to AI have gained traction as they have proved to be successful at important tasks for speech recognition and natural language
Andrew S. Gordon
[email protected] Rob Miller [email protected] Leora Morgenstern [email protected] Gy¨orgy Tur´an [email protected] 1
University of Southern California, Los Angeles, CA, USA
2
University College London, London, UK
3
Palo Alto Research Center, Palo Alto, CA, USA
4
University of Illinois at Chicago, Chicago, IL, USA
5
University of Szeged, Szeged, Hungary
A.S. Gordon et al.
understanding [10, 11]. Over the last decade, neural network techniques in particular have shown great promise in significantly increasing the level of performance not only on such tasks, but also on previously intractable problems such as image recognition [12] and those involving complex planning, such as playing Go [13]. More recently, neural nets have succeeded beyond expectations on challenges such as the Winograd Schema Challenge [14], which was specifically designed to assess whether a system has commonsense knowledge and can perform commonsense reasoning. This success has been achieved despite the lack of large formalized commonsense knowledge bases or demonstrated ability of software systems to perform commonsense reasoning [15]. Yet neural network techniques are not a panacea. They are vulnerable to adversarial attacks [16], magnify biases in training data, and are often prone to the sorts of errors that would be immediately obvious to a human and that indicate a serious lack of common sense [17]. Formal approaches to commonsense knowle
