Intentional Forgetting in Artificial Intelligence Systems: Perspectives and Challenges

Abstract

Current trends, like digital transformation and ubiquitous computing, yield in massive increase in available data and information. In artificial intelligence (AI) systems, capacity of knowledge bases is limited due to computational complexity of many inference algorithms. Consequently, continuously sampling information and unfiltered storing in knowledge bases does not seem to be a promising or even feasible strategy. In human evolution, learning and forgetting have evolved as advantageous strategies for coping with available information by adding new knowledge to and removing irrelevant information from the human memory. Learning has been adopted in AI systems in various algorithms and applications. Forgetting, however, especially intentional forgetting, has not been sufficiently considered, yet. Thus, the objective of this pa- per is to discuss intentional forgetting in the context of AI systems as a first step. Starting with the new priority research program on ‘Inten- tional Forgetting’ (DFG-SPP 1921), definitions and interpretations of intentional forgetting in AI systems from different perspectives (knowl- edge representation, cognition, ontologies, reasoning, machine learning, self-organization, and distributed AI) are presented and opportunities as well as challenges are derived.

Full text

Intentional Forgetting in Artificial Intelligence
Systems: Perspectives and Challenges
Ingo J. Timm1, Steffen Staab2, Michael Siebers3, Claudia Schon2,
Ute Schmid3, Kai Sauerwald4, Lukas Reuter1, Marco Ragni5,
Claudia Nieder´ee6, Heiko Maus7, Gabriele Kern-Isberner8, Christian Jilek7,
Paulina Friemann5, Thomas Eiter9, Andreas Dengel7, Hannah Dames5,
Tanja Bock8, Jan Ole Berndt1, Christoph Beierle4∗
1Trier University {itimm,reuter,berndt}@uni-trier.de
2University Koblenz-Landau {staab,schon}@uni-koblenz.de
3University of Bamberg {michael.siebers,ute.schmid}@uni-bamberg.de
4FernUniversit¨at in Hagen
{kai.sauerwald,christoph.beierle}@fernuni-hagen.de
5Albert-Ludwigs-Universit¨at Freiburg
{ragni,friemanp,damesh}@cs.uni-freiburg.de
6L3S Research Center Hannover niederee@l3s.de
7German Research Center for Artificial Intelligence (DFKI)
{christian.jilek,andreas.dengel,heiko.maus}@dfki.de
8TU Dortmund
tanja.bock@tu-dortmund.de,gabriele.kern-isberner@cs.uni-dortmund.de
9TU Wien eiter@kr.tuwien.at
Abstract. Current trends, like digital transformation and ubiquitous
computing, yield in massive increase in available data and information.
In artificial intelligence (AI) systems, capacity of knowledge bases is
limited due to computational complexity of many inference algorithms.
Consequently, continuously sampling information and unfiltered storing
in knowledge bases does not seem to be a promising or even feasible
strategy. In human evolution, learning and forgetting have evolved as
advantageous strategies for coping with available information by adding
new knowledge to and removing irrelevant information from the human
memory. Learning has been adopted in AI systems in various algorithms
and applications. Forgetting, however, especially intentional forgetting,
has not been sufficiently considered, yet. Thus, the objective of this pa-
per is to discuss intentional forgetting in the context of AI systems as
a first step. Starting with the new priority research program on ‘Inten-
tional Forgetting’ (DFG-SPP 1921), definitions and interpretations of
intentional forgetting in AI systems from different perspectives (knowl-
edge representation, cognition, ontologies, reasoning, machine learning,
self-organization, and distributed AI) are presented and opportunities as
well as challenges are derived.
Keywords: Artificial Intelligence Systems ·Capacity and Efficiency of
Knowledge-Based Systems ·(Intentional) Forgetting.
∗Author names appear in alphabetical order

2 Timm et al.
1 Introduction
Today’s enterprises are dealing with massively increasing digitally available data
and information. Current technological trends, e.g., Big Data, focus on aggre-
gation, association, and correlation of data as a strategy to handle information
overload in decision processes. From a psychological perspective, humans are cop-
ing with information overload by selective forgetting of knowledge. Forgetting
can be defined as non-availability of a previously known certain piece of infor-
mation in a specific situation [29]. It is an adaptive function to delete, override,
suppress, or sort out outdated information [4]. Thus, forgetting is a promising
concept of coping with information overload in organizational contexts.
The need for forgetting has already been recognized in computer science
[17]. In logics, context-free forgetting operators have been proposed, e.g., [6,30].
While logical forgetting explicitly modifies the knowledge base (KB), various
machine learning approaches implicitly forget details by abstracting from their
input data. In contrast to logical forgetting, machine learning can be used to
reduce complexity by aggregating knowledge instead of changing the size of a
KB. As a third approach, distributed AI (DAI) focuses on reducing complexity by
distributing knowledge across agents [21]. These agents ’forget’ at the individual
level while the overall system ’remembers’ through their interaction.
For humans, forgetting is also an intentional mechanism to support decision-
making by focusing on relevant knowledge [4,22]. Consequently, the questions
arise when and how humans can intentionally forget and when and how intel-
ligent systems should execute forgetting functions. The new priority research
program on ”Intentional Forgetting in Organizations” (DFG-SPP 1921) has been
initiated to elaborate an interdisciplinary paradigm. Within the program, re-
searchers from computer science and psychology are interdisciplinarily collabo-
rating on different aspects of intentional forgetting in eight tandem projects.1
With a strong focus (five projects) on AI systems, multiple perspectives are
researched ranging from knowledge representation, cognition, ontologies, rea-
soning, machine learning, self-organization, and DAI. In this paper we bring
together these perspectives as a first building block for establishing a common
understanding of intentional forgetting in AI. Contributions of this paper are
the identification of AI research fields and their challenges.
2 Knowledge Representation and Cognition: FADE
The goal of FADE (Forgetting through Activation, reDuction and Elimination)
is to support the effortful preselection and aggregation of information in in-
formation flows, leading to a reduction of the user’s workload, by integrating
methods from cognitive and computer science: Knowledge structures in orga-
nizations and mathematical and psychological modeling approaches of human
memory structures in cognitive architectures are analyzed. Functions for pri-
orization and forgetting that may help to compress and reduce the increasing
1http://www.spp1921.de/projekte/index.html.de

Intentional Forgetting in Artificial Intelligence Systems 3
amount of data are designed. Furthermore, a cognitive computational system
for forgetting is developed that offers the opportunity to determine and adapt
system model parameters systematically and makes them transparent for every
single knowledge structure. This model for forgetting is evaluated for its fit to a
lean workflow and readjusted in the context of the ITMC of the TU Dortmund.
While forgetting is often attributed negatively in everyday life, forgetting
can offer an effective and beneficial reduction process to allow humans to focus
on information of higher relevance. Features of the cognitive forgetting process
which are crucial to the FADE project work are that information never gets lost
but instead has a level of activation [1], and that the relevance of information
depends on its connection to other information and its past usage. Moreover,
information characteristics require different forms of forgetting; in particular,
insights from knowledge representation and reasoning can help to further re-
fine declarative knowledge, and differentiate between assertional knowledge and
conceptual knowledge. Finally, it can be expected that cognitive adequacy of for-
getting approaches will improve the human-computer interaction significantly.
The project FADE focusses on formal methods that are apt to model the
epistemic and subjective aspects of forgetting [3,13]. Here, the wide variety of
formalisms of nonmonotonic reasoning and belief revision are extremely helpful
[2]. The challenge is to adapt these approaches to model human-like forgetting,
and to make them usable in the context of organizations. As a further mile-
stone, these adapted formal methods are integrated into cognitive architectures
providing a formal-cognitive frame for forgetting operations [23,24].
3 Ontologies and Reasoning: EVOWIPE
New products are often developed by modifying the model of an already existing
product. Assuming that large parts of the product model are represented in a
KB, the EVOWIPE project supports this reuse of existing product models by
providing methods to intentionally forget aspects from a KB that are not appli-
cable to the new product [14]. E.g., the major part of the product model of the
VW e-Golf (with electric motor) is based on the concept of the VW Golf with
combustion engine. However, (i) changes, (ii) additions and (iii) forgetting ele-
ments of the original product model are necesarry, e.g. (i) connecting the engine,
(ii) adding a temperature control system for the batteries, and (iii) forgetting the
fuel tank, fuel line and exhaust gas treatment. EVOWIPE aims at developing
methods to support the product developer in the process of forgetting aspects
from product models represented in KBs by developing the following operators
for intentional forgetting: Forgetting of inferred knowledge, restoring forgotten
elements, temporary forgetting, representation of place markers in forgetting,
cascading forgetting.
These operators bear similarities to deletion operators known in knowledge
representation (cf. Section 2). Indeed, we represent knowledge about product
models by transforming existing product model data structures into an OWL-
based representation and build on existing research that accesses such KBs using

4 Timm et al.
SPARQL update queries. These queries allow not only for deleting knowledge
but also for inserting new knowledge. Therefore, the interplay of deletion and
insertion is investigated in the project as well [25]. To accomplish cascading for-
getting, dependencies occurring in the KB have to be specified. They can be
added as metaproperties into the KB [10]. These dependencies can be added
manually, however the project partners are currently working on methods to au-
tomatically extract dependencies from the product model. Dependency-guided
semantics for SPARQL update queries use these dependencies to accomplish
the desired cascading behavior described above [15]. By developing these opera-
tors, the EVOWIPE project extends the product development process to include
stringent methods for intentional forgetting, ensuring that the complexity inher-
ent in the product model, the product development process and the forgetting
process itself can be mastered by the product developer.
4 Machine Learning: Dare2Del
Dare2Del is a system designed as context-aware cognitive companion [9,26] to
support forgetting of digital objects. The companion will help users to delete or
archive digital objects which are classified as irrelevant and it will support users
to focus on a current task by fading-out or hiding digital information which is
irrelevant in a given task context. In collaboration with psychology, it is investi-
gated for which persons and in which situations information hiding can improve
task performance and how explanations can establish trust of users in system
decisions. The companion is based on inductive logic programming (ILP) [18] – a
white-box machine learning approach based on Prolog. ILP allows learning from
small sets of training data, a natural combination of reasoning and learning, and
the incorporation of background knowledge. ILP has been shown to be able to
provide human-understandable classifiers [19].
For Dare2Del to be a cognitive companion, it should be able to explain system
decisions to users and be adaptive. Therefore, we currently design an incremental
variant of ILP to allow for interactive learning [8]. Dare2Del will take into ac-
count explanations given by the user. E.g., if a user decides that an object should
not be deleted, he or she can select one or more predicates (presented in natural
language) which hold for the object and which are the reason why it should
not be deleted. Subsequently, Dare2Del has to adapt its model. As application
scenarios for Dare2Del we consider administration as well as connected industry.
In the context of administration, users will be supported to delete irrelevant files
and Dare2Del will help to focus attention by hiding irrelevant columns in tables.
In the context of connected industry, quality engineers are supported in iden-
tifying irrelevant measurements and irrelevant data for deletion. Alternatively,
measurements and data can be hidden in the context of a given control task. We
believe that Dare2Del can be a helpful companion to relieve humans from the
cognitive burden of complex decision making which is often involved when we
have to decide whether some digital object will be relevant in the future or not.

Intentional Forgetting in Artificial Intelligence Systems 5
5 Self-Organization: Managed Forgetting
We investigate intentional forgetting in grass-roots (i.e. decentralized and self-
organizing) organizational memory, where knowledge acquisition is incorporated
into daily activities of knowledge workers. In line with this, we have introduced
Managed Forgetting (MF) [20] - an evidence-based form of intentional forgetting,
where no explicated will is required: what to forget and what to focus on is
learned in a self-organizing and decentralized way based on observed evidences.
We consider two forms of MF: memory buoyancy empowering forgetful infor-
mation access and context-based inhibition easing context switches. We apply MF
in the Semantic Desktop, which semantically links information items in a ma-
chine understandable way based on a Personal Information Model (PIMO) [11].
Shared parts of individual PIMOs form a basis for an Organizational Memory.
As a key concept for this form of MF we have presented Memory Buoy-
ancy (MB) [20], which represents an information item’s current value for the
user. It follows the metaphor of less relevant items “sinking away” from the
user, while important ones are pushed closer. MB value computation has been
investigated for different types of resources [5,28] and is based on a variety of
evidences (e.g. user activities), activation propagation as well as on heuristics.
MB values provide the basis for forgetful access methods such as hiding or
condensation [11], adaptive synchronization and deletion, and forgetful search.
Most knowledge workers experience frequent context switches due to mul-
titasking. Other than the gradual changes of MB in the first form of MF, in
the case of context switches, changes are far more abrupt. We, therefore, be-
lieve that approaches based on the concept of inhibition [16], which temporarily
hide resources of other contexts could be employed here, e.g. in a kind of self-
tidying and self-(re)organizing context spaces [12]. Our current research focuses
on combining both forms of MF.
6 Distributed Artificial Intelligence: AdaptPRO
In DAI, (intelligent) agents encapsulate knowledge which is deeply connected to
domain, tasks and action [21]. They are intended to perceive their environment,
react to changes, and act autonomously by (social) deliberation. Forgetting is
implicitly a subject of research, e.g., Belief Revision (cf. section 2) or possible-
worlds semantics [31]. By contrast, the team perspective of forgetting, i.e., change
of knowledge distribution, roles, and processes have not been analyzed yet.
In AdaptPRO, we focus on these aspects by adopting intentional forgetting
in teams from psychology. We define intentional forgetting as the reorganization
of knowledge in teams. The organization of human team knowledge is known as
team cognition (TC). TC describes the structure in which knowledge is mentally
represented, distributed, and anticipated by members to execute actions [7]. The
concept of TC can be used to model knowledge distribution in agent systems
as well. In terms of knowledge distributions, organization of roles and processes
are implemented by allocating, sharing or dividing knowledge. If certain team

6 Timm et al.
members are specialized on particular areas, other agents can ignore information
related to this area [27]. Especially, when cooperating, it is important for agents
to share their knowledge about task- and team-relevant information. Particu-
larly in case of disturbances, redundant knowledge and task competences enable
robust teamwork. To strike a balance between sharing and dividing knowledge,
i.e., efficient and robust teamwork, AdaptPRO applies an interdisciplinary ap-
proach of modeling, analyzing and adapting knowledge structures in teams and
measure their implications on individual and team perspective.
7 Challenges and Future Work
We have presented perspectives on intentional forgetting in AI systems. Their
key opportunities can be summarized as follows: (a) Establishing guidelines that
help to implement human-like forgetting for organizations by bridging Cognition
and Organizations with formal AI methods. (b) Mastering information overload
by (temporary) forgetting and restoring of knowledge with respect to inferred
and cascading knowledge structures. (c) Supporting decision-making of humans
by forgetting digital objects with comprehensive knowledge management and
machine learning. (d) Assisting organizational knowledge management with in-
tentional forgetting by self-organization and self-tidying. (e) Adapting processes
and roles in organizations by reorganization of knowledge distribution. In or-
der to tap into these opportunities, the following challenges must be overcome:
(1) Merge concepts of (intentional) forgetting in AI in a common terminology.
(2) Formalize kinds of knowledge and forgetting to make prerequisites and aims
of forgetting operations transparent and study their formal properties. (3) Inves-
tigate whether different forms of knowledge require different techniques of forget-
ting. (4) Accomplish efficient remembering of knowledge. (5) Develop temporar-
ily forgetting information from a KB. (6) Develop of an incremental probabilistic
approach to inductive logic programming which allows interactive learning by
mutual explanations. (7) Generate helpful explanations in form of verbal jus-
tifications and by providing examples or counterexamples. (8) Develop correct
interpretation on user activities, work environment, and information to initi-
ate appropriate forgetting measures. (9) Characterize knowledge in teams and
DAI-Systems and develop formal operators for reallocating, extending, and for-
getting information.
These challenges foster an important basis for AI research in the next years.
Furthermore, intentional forgetting has the potential to evolve to a mandatory
function of next generation AI systems, which become capable of coping with
our days’ complexity and data availability.
Acknowledgments. The authors are indebted to the DFG for funding this
research: Dare2Del (SCHM1239/10-1), EVOWIPE (STA572/15-1), FADE (BE
1700/9-1, KE1413/10-1, RA1934/5-1), Managed Forgetting (DE420/19-1, NI1760
1-1), and AdaptPro (TI548/5-1). We would also like to thank our project part-
ners for their fruitful discussion: C. Antoni, T. Ellwart, M. Feuerbach, C. Frings,
K. G¨obel, P. K¨ugler, C. Niessen, Y. Runge, T. Tempel, A. Ulfert, S. Wartzack.

Intentional Forgetting in Artificial Intelligence Systems 7
References
1. Anderson, J.R.: How can the human mind occur in the physical universe? Oxford
University Press, New York (2007)
2. Beierle, C., Kern-Isberner, G.: Semantical investigations into nonmonotonic and
probabilistic logics. Annals of Mathematics and Artificial Intelligence 65(2), 123–
158 (2012)
3. Beierle, C., Eichhorn, C., Kern-Isberner, G.: Skeptical inference based on
c-representations and its characterization as a constraint satisfaction prob-
lem. In: Foundations of Information and Knowledge Systems - 9th Interna-
tional Symposium, FoIKS 2016, Linz, Austria, March 7-11, 2016. Proceedings.
Lecture Notes in Computer Science, vol. 9616, pp. 65–82. Springer (2016).
https://doi.org/10.1007/978-3-319-30024-5 4
4. Bjork, E.L., Anderson, M.C.: Varieties of goal-directed forgetting. In: Golding,
J.M., MacLeod, C. (eds.) Intentional Forgetting: Interdisciplinary Approaches, pp.
103–137. Lawrence Erlbaum: Mahwah, NJ (1998)
5. Ceroni, A., Solachidis, V., Nieder´ee, C., Papadopoulou, O., Kanhabua, N., Mezaris,
V.: To keep or not to keep: An expectation-oriented photo selection method for
personal photo collections. In: Proc. of the 5th ACM Int’l Conf. on Multimedia
Retrieval, Shanghai, China, June 23-26, 2015. pp. 187–194 (2015)
6. Delgrande, J.P.: A knowledge level account of forgetting. J. Artif. Intell. Res. 60,
1165–1213 (2017)
7. Ellwart, T., Antoni, C.H.: Shared and distributed team cognition and informa-
tion overload. evidence and approaches for team adaptation. In: Marques, R.P.F.,
Batista, J.C.L. (eds.) Information and Communication Overload in the Digital
Age, pp. 223–245. IGI Global, Hershey (2017)
8. Fails, J.A., Olsen Jr, D.R.: Interactive machine learning. In: Proceedings of the 8th
International Conference on Intelligent User Interfaces. pp. 39–45. ACM (2003)
9. Forbus, K.D., Hinrichs, T.R.: Companion cognitive systems: A step toward human-
level AI. AI magazine 27(2), 83 (2006)
10. Guarino, N., Welty, C.A.: An overview of ontoclean. In: Handbook on Ontologies,
pp. 201–220. International Handbooks on Information Systems, Springer (2009)
11. Jilek, C., Maus, H., Schwarz, S., Dengel, A.: Diary generation from personal in-
formation models to support contextual remembering and reminiscence. In: 2015
IEEE Int’l Conf. on Multimedia & Expo Workshops, ICMEW 2015. pp. 1–6 (2015)
12. Jilek, C., Schr¨oder, M., Schwarz, S., Maus, H., Dengel, A.: Context spaces as the
cornerstone of a near-transparent & self-reorganizing semantic desktop. In: The
Semantic Web: ESWC 2018 Satellite Events. Springer (2018)
13. Kern-Isberner, G., Bock, T., Sauerwald, K., Beierle, C.: Iterated contraction of
propositions and conditionals under the principle of conditional preservation. In:
Benzm¨uller, C., Lisetti, C.L., Theobald, M. (eds.) GCAI 2017, 3rd Global Confer-
ence on Artificial Intelligence, Miami, FL, USA, 18-22 October 2017. Proceed-
ings. EPiC Series in Computing, vol. 50, pp. 78–92. EasyChair (2017), http:
//www.easychair.org/publications/paper/DTmX
14. Kestel, P., Luft, T., Schon, C., K¨ugler, P., Bayer, T., Schleich, B., Staab, S.,
Wartzack, S.: Konzept zur zielgerichteten, ontologiebasierten Wiederverwendung
von Produktmodellen. In: Krause, D., Paetzold, K., S. Wartzack, S. (eds.) Design
for X. Beitr¨age zum 28. DfX-Symposium. pp. 241–252. TuTech Verlag, Hamburg
(2017)

8 Timm et al.
15. K¨ugler, P., Kestel, P., Schon, C., Marian, M., Schleich, B., Staab, S., Wartzack, S.:
Ontology-based approach for the use of intentional forgetting in product develop-
ment. In: DESIGN Conference Dubrovnik (2018)
16. Levy, B.J., Anderson, M.C.: Inhibitory processes and the control of memory re-
trieval. Trends in cognitive sciences 6(7), 299–305 (2002)
17. Markovitch, S., Scott, P.D.: Information filtering: Selection mechanisms in learning
systems. Machine Learning 10(2), 113–151 (Feb 1993)
18. Muggleton, S., De Raedt, L.: Inductive logic programming: Theory and methods.
The Journal of Logic Programming 19, 629–679 (1994)
19. Muggleton, S.H., Schmid, U., Zeller, C., Tamaddoni-Nezhad, A., Besold, T.: Ultra-
strong machine learning-comprehensibility of programs learned with ilp. Machine
Learning (2018)
20. Nieder´ee, C., Kanhabua, N., Gallo, F., Logie, R.H.: Forgetful digital memory:
Towards brain-inspired long-term data and information management. SIGMOD
Record 44(2), 41–46 (2015)
21. O’Hare, G.M.P., Jennings, N.R. (eds.): Foundations of Distributed Artificial Intel-
ligence. John Wiley & Sons, Inc., New York (1996)
22. Payne, B.K., Corrigan, E.: Emotional constraints on intentional forgetting. Journal
of Experimental Social Psychology 43(5), 780–786 (2007)
23. Ragni, M., Sauerwald, K., Bock, T., Kern-Isberner, G., Friemann, P., Beierle, C.:
Towards a formal foundation of cognitive architectures. In: Proceedings of the 40th
Annual Meeting of the Cognitive Science Society, CogSci 2018, Madison, US, 25-28
July 2018 (2018), (to appear)
24. Sauerwald, K., Ragni, M., Bock, T., Kern-Isberner, G., Beierle, C.: On a formal-
ization of cognitive architectures. In: Proceedings of the 14th Biannual Conference
of the German Cognitive Science Society. Darmstadt (2018), (to appear)
25. Schon, C., Staab, S.: Towards SPARQL instance-level update in the presence of
OWL-DL tboxes. In: JOWO. CEUR Workshop Proceedings, vol. 2050. CEUR-
WS.org (2017)
26. Siebers, M., G¨obel, K., Niessen, C., Schmid, U.: Requirements for a companion
system to support identifying irrelevancy. In: International Conference on Com-
panion Technology, ICCT 2017, Ulm, Germany, September 11-13, 2017. pp. 1–2.
IEEE (2017)
27. Timm, I.J., Berndt, J.O., Reuter, L., Ellwart, T., Antoni, C., Ulfert, A.S.: To-
wards multiagent-based simulation of knowledge management in teams. In: Leyer,
M., Richter, A., Vodanovich, S. (eds.) Flexible knowledge practices and the Dig-
ital Workplace (FKPDW). Workshop within the 9th Conference on Professional
Knowledge Management. pp. 25–40. KIT: Karlsruhe (2017)
28. Tran, T., Schwarz, S., Nieder´ee, C., Maus, H., Kanhabua, N.: The forgotten needle
in my collections: Task-aware ranking of documents in semantic information space.
In: CHIIR-16. ACM Press (2016)
29. Tulving, E.: Cue-dependent forgetting: When we forget something we once knew,
it does not necessarily mean that the memory trace has been lost; it may only be
inaccessible. American Scientist 62(1), 74–82 (1974)
30. Wang, Z., Wang, K., Topor, R., Pan, J.Z.: Forgetting for knowledge bases in dl-lite.
Annals of Mathematics and Artificial Intelligence 58(1), 117–151 (Feb 2010)
31. Werner, E.: Logical foundations of distributed artificial intelligence. pp. 57–117.
John Wiley & Sons, Inc., New York (1996)