Intentional Forgetting: An Emerging Field in AI and Beyond

Christoph Beierle, Ingo J. Timm
2018 Künstliche Intelligenz  
In human evolution, learning has evolved as a key strategy for adaptation. While animals adapt on the basis of their own experiences, humans integrate complex forms of interaction and information exchange in the learning process to include other people's experiences in addition to stimulus-responseconditioning. By lifting parts of the adaptation to the cognitive level, humans consciously reason about the learning process, i.e., they continuously perceive new information and extend their
more » ... e by learning. Due to limited cognitive capacity, new information is not just added but irrelevant information is sorted out, e.g., we know what we have had for breakfast today but fade out breakfast details of the last days. We have established memory structures enabeling us to remember or derive knowledge fragments with considerable success, albeit possibly requiring also considerable effort, even when we have forgotten about them. Human learning and forgetting is not just steered by subcognitive routines, as in the example of remembering breakfast, but seems also to rely on a memory organization supporting to intentionally deliberate on previously or implicitly known information fragments. In current research both in psychology and information systems, intentionally balancing learning and forgetting is of great interest to meet information overload as a new challenge in the information or digital society. Current trends, like the digital transformation and ubiquitous computing, cause a massive increase in available data and information. However, the problem of information overload does not only tackle human but also computer systems being bounded by physical memory. From a complexity perspective, many (if not most) AI algorithms exhibit exponential runtime or space complexity. As a result, even slight increases in the amount of information available may have a disproportionate effect on the runtime or memory requirements for those algorithms. Therefore, the need for forgetting in computer science has already been recognized in academics and even in politics. From a normative perspective, the question arises whether there is a right to be forgotten, e.g., whether it is mandatory to provide mechanisms in the Internet or online social networks, such that data or information that has become available once can be removed or forgotten later on [6] . From a technological perspective, this interpretation of forgetting addresses technical mechanisms for deleting or reordering data and information. While this is an inherent functionality of any computer system, in cloud and cluster computing it remains a severe challenge as for an individual living in todays digital epoch keeping control of the specific physical location of data becomes practically infeasible. Should AI Systems Forget? Especially in AI systems, a sophisticated concept of forgetting is required, as knowledge and the acquisition of knowledge is ubiquitous: "Although additional knowledge can enhance the performance of a problem solver by reducing its requirements, it can also have exactly the opposite effect. In particular, additional knowledge can have the effect of greatly increasing search time [...]" [9, p. 116]. For knowledge-based systems or in the context of semantic technologies, there is a dominance of logic-based representation of knowledge. To implement such systems, operators for adding new knowledge, e.g., facts or rules, are provided. Depending on the specific logic, there are prerequisites to be considered when manipulating the knowledge base, e.g., consistency. Forgetting has also been addressed in logics; for instance, forgetting operators have been proposed
doi:10.1007/s13218-018-00574-x fatcat:jv5mzmx7irf2zlhrjeqzlkg7k4