Institut für Kognitionswissenschaft

Institute of Cognitive Science

Navigation und Suche der Universität Osnabrück



Some topics for BA and MA theses


Methods for artificial neural network explicability


A current challange in Artificial Intelligence and Machine Learning is to find methodologies on how to decode representations of problems within the weight-space of artificial neural networks (ANNs). while some approaches focus on artificial neural networks trained on visual data to enhance interpretability of activity through relevance propagation, others specify formal models for aquiring insight into semantic representations in network topology.

Within the interdisciplinary scope between Neuroscience, Artificial Intelligence and mathematics, students are encouraged to pick from the following Thesis topics on ANN explicability:

  • Perform literature research on and implementation of approaches already explored in previous studies.
  • Conceptualizing and implementing novell methods derived from neuroscience and other interdisciplinary fields related to Cognitive Science.
  • Developing formal theories on weight space analysis.
Contact: Johannes Schrumpf

Analyzing Twitter Discussions with Weighted Bipolar Argumentation Frameworks

  Recently, computational argumentation has been investigated as a tool for analyzing discussions on social media platforms. For instance, a Twitter discussion can be modeled as a graph where each node denotes a tweet and an edge from tweet t1 to tweet t2 indicates that t1 replies to t2. Sentiment analysis tools can be applied to classify edges as attacks (tweets criticizing other tweets), supports (tweets supporting other tweets) or other (currently irrelevant relations). A social relevance score can be assigned to tweets based on statistical data like the number of followers of the author or the number of retweets.

Given this graph with social relevance scores for each tweet and attack and support relations between tweets, we are interested in assigning a strength value between 0 and 1 to each position. 0 means that the position should be rejected and 1 means that it should be accepted. A value of 0.5 expresses indifference, whereas values close to 0 (1) express that we tend to reject (accept) the position. If there are no attackers and supporters, the strength should be just the social relevance score. If attackers and supporters are present, the strength must be adapted based on the strength of attackers and supporters. During the last years, some weighted bipolar argumentation frameworks have been presented that can compute such strength values.

In this thesis, weighted bipolar argumentation frameworks shall be evaluated for the analysis of Twitter discussions. To this end, you should build up on existing work in this field. In [1], a related framework has been presented that can only accept or reject arguments (no graded acceptance values between 0 and 1) and considers only attack relations (no support). However, the ideas to compute social relevance scores and to identify attacking tweets using sentiment analysis tools will also be helpful for this work. The frameworks that should be considered in this thesis are the DF-QuAD algorithm from [2], the Euler-based restricted semantics from [3] and another currently unpublished framework that has been developed in our group.

In order to work on this thesis successfully, you should have basic knowledge of machine learning and artificial intelligence in general (Introduction to AI and Methods of AI will be sufficient), should be comfortable with formal approaches and have proficient programming skills in Java or Python.

[1] Alsinet, T., Argelich, J., Béjar, R., Planes, J., Cemeli, J., & Sanahuja, C. 2017. A Distributed Approach for the Analysis of Discussions in Twitter. In Proceedings of the 3rd International Workshop on Social Influence Analysis (SocInf 2017), 45-56.

[2] Rago, A.; Toni, F.; Aurisicchio, M.; and Baroni, P. 2016. Discontinuity-free decision support with quantitative argumentation debates. In International Conference on Principles of Knowledge Representation and Reasoning (KR), 63–73.

[3] Amgoud, L., and Ben-Naim, J. 2017. Evaluation of arguments in weighted bipolar graphs. In European Conference on Symbolic and Quantitative Approaches to Reasoning and Uncertainty, 25–35.

Contact: Nico Potyka

Computational Storytelling using Multi-Agent Simulation


To model computational creativity in the domain of stories, one first needs to develop a computational representation for these multi-faceted and dynamic works of ingenuity. We approach this by representing a narrative's story-world and characters as a multi-agent simulation system. The plot of a story, in this case, emerges from the (inter-) actions of the system's agents in combination with the internal events in the agents' reasoning cycles: the plans, beliefs, desires and emotions that make them act and react in certain ways.
Our system can be extended in several ways, which are all potential thesis projects:

  • A narrative theory based measure of the quality of the plot called tellability needs to be computed based on a graph representation of the produced plots, and its validity needs to be empirically evaluated.
  • The employed character architecture includes a cognitively inspired model of personality traits and affects, which can be used to explore a space of possible plots. Yet it remains to empirically study whether different personality parameters result in a reader-impression of different personalities in characters, and whether readers can correctly derive character's moods and emotions from their actions in the plot.
  • Fictional characters are social (paper) beings, they constantly try to outwit, deceive or simply understand each other. An important property of fictional minds should thus be a mind reading ability, a "Theory of Mind". Implementing a domain-independent ToM on the architectural level of the system is thus important to generate most interesting tales. Achieving this through a connection of planning and recursive simulations could be explored by a challenging MSc thesis.
  • Many stories serve the purpose of representing a characters mental and social maturation process. This means that (some) characters need to be able to change through the course of a plot. Learning and character development are two phenomena that could be integrated into the storytelling system, to allow it to model a broader range of narratives. Such a thesis would need to perform narratological, theoretical research as well as practical implementation work and is better suited as a MSc thesis.
Contact: open

Conceptual Spaces for Artificial Intelligence


The cognitively inspired framework of conceptual spaces proposes to encode knowledge based on geometric representations: Objects are mapped onto points in a semantic feature space and concepts are mapped onto convex regions in this space. This approach has various advantages:

  • One can connect abstract symbols used in classic AI systems (represented as concepts) to lower-level semantic features (represented as dimensions of the space).
  • Machine learning algorithms can be readily applied, as they typically also make use of a feature space.
  • As the individual dimensions of the conceptual space are cognitively meaningful, the resulting regions can be interpreted.
  • The whole approach is grounded in psychological findings, making it cognitively plausible.

In the context of our current research on conceptual spaces for AI, we have already developed and implemented a thorough mathematical formalization of the framework. The following potential thesis topics explore applications and extensions of this formalization:

  • There are different types of reasoning in conceptual spaces, all of them based on a geometric approach. The goal of a thesis in this area is to algorithmically define, implement, and evaluate these reasoning techniques in the context of our formalization. Potential subtopics include:
    • Concept combination: Combining existing conceptual regions in order to derive the meaning of compound terms like “small green apple”.
    • Similarity-based reasoning: Using conceptual similarity to implement common sense reasoning strategies – if you like oranges, then you might also like mandarins.
    • Interpolative reasoning: Using conceptual betweenness to implement common sense reasoning strategies – if both bachelor and PhD students have to pay an enrollment fee, this probably also applies to master students.
  • Potentially, existing concepts in a given conceptual space can scaffold learning from very limited data. In one-shot learning, a new conceptual region is learned based on a single example. This can potentially be achieved by generalizing from existing conceptual regions that partially fit the given example. In zero-shot learning, a new conceptual region is learned without any examples, but based on an abstract description of the concept (e.g., “oranges are sweet, round, orange and weigh about 130 g”). A thesis in this area defines, implements, and evaluates an algorithm for solving these learning problems in the context of our formalization.
  • A huge fly is a lot smaller than a tiny house. The meaning of relative properties like “huge” and “tiny” depends on the contrast class it is being applied to (i.e., “fly” or “house”). The goal of this thesis project is to extend our formalization such that it can also represent such relative properties.
  • Conceptual knowledge always includes information about part-whole relations (e.g., a car has four wheels). There has already been a proposal about how to represent such relations in conceptual spaces, however this proposal has never been implemented. The goal of this thesis project is to incorporate this proposal into our formalization, to implement it, and to investigate whether it is feasible and useful from a practical point of view. One can also investigate whether this can generalize to more kinds of relations.

[All topics are available as both bachelor and master thesis.]

Contact: Lucas Bechberger

(Deep) Learning in the Health Domain


Through a cooperation with a health consulting company, we have access to a large data set of health data including both structured information (hierarchical codes for diagnosed illnesses and applied treatments) and unstructured information (laboratory values for blood samples). This data set is well suited for exploratory studies on using machine learning in the health domain to predict probable diagnoses, how long the patient will stay in the hospital, what kind of treatment is appropriate, etc.

Dialog Systems for Cognitive Robotics


Currently ongoing work focuses on the usage of a Pepper robot as a tutor in a classroom setting. In order to allow for an interactive dialog with students, we are interested in developing a dialog system in this cognitive robotics context.

Learning and User Modeling for E-Learning Platforms



Heuristics for theory projection in analogical reasoning


Heuristic-Driven Theory Projection (HDTP) is a framework for analogical reasoning currently under development at the IKW. It works in two stages by first identifying common structures in the modelling of two different domains and then transferring knowledge und proposing new conclusions.
Several steps in this process are based on heuristic knowledge. The thesis shall propose and implement different heuristics and examine their value in process of analogy making.

The semantics of theory projection


Heuristic-Driven Theory Projection (HDTP) is based on an extended form of anti-unification of logical theories. It introduces certain second-order elements which seem to be well-founded on a syntactic level but whose semantics needs to be further investigated. A thesis should develop a model theoretic semantics and show its applicability in the description of analogies.

Comparing analogy models


Heuristic-Driven Theory Projection (HDTP) is a framework for analogical reasoning currently under development at the IKW. The representation is essentially logic-based. HDTP works in two stages by first identifying common structures in the modelling of two different domains and then transferring knowledge und proposing new conclusions.
Besides HDTP exist many other approaches to analogical mapping which use different representational formalisms and have different underlying assumptions how analogies occur. E.g., according to the Structure Mapping Theory (SMT) by Gentner, a domain is represented as a structure (a graph) with objects, attributes and relations. Analogies are identified via structural commonalities. The Analogical Constraint Mapping Engine (ACME) by Holyoack and Thagard uses neuronal networks to identify mappings. In turn, Indurkhya and Dastani use an algebraic framework for representation of domains.
The thesis shall select a number of approaches and compare them against HDTP with respect to their analogical mapping and inference mechanisms.

Analogical reasoning and learning


Analogy making is the process by which humans extract operators used to solve one problem and map them onto a solution for another problem. The thesis shall investigate experimentally the human way of analogical reasoning and learning.
The findings can be compared to other learning strategies (e.g. inductive, abductive and deductive reasoning) or they can be tested against the analogical inference mechanisms in HDTP.

Learning first-order logical theories with neural networks in reasoning processes


The well-known gap in Cognitive Science and Artificial Intelligence between symbolic and subsymbolic modelings seems to be a hard problem. An approach developed at the IKW to bridge this gap in one direction uses a translation of logical theories to a semisymbolic representation, namely to so-called Topos theory. This representation is homogeneous, variable-free, and uses only one inherent operation (concatenation of arrows). The semisymbolic level can be used to train a feedforward neural network that is not only learning the logical input theory, but rather the logical closure of this theory, i.e. a model of the underlying theory.
The thesis shall develop this approach further by applying it to complex logical input theories in reasoning and planning domains.

Comparing algebraic frameworks used in AI


Rule-based and procedural approaches are dominating the field of reasoning in artificial intelligence. Examples are methods used for inductive, case-based, or qualitative reasoning. Alternatively, algebraic approaches can be used. Such approaches focus on structural representations and provide a powerful and formally sound foundation for inference algorithms. Examples are the usage of anti-unification in analogy and induction or relational algebras for qualitative reasoning.
The thesis shall compare different algebraic methods with respect to their expressive strength and other theoretic features.

Assessment of solutions for programming assignments in PROLOG and LISP


Assessing assignments and exams produces a big workload for lecturers and tutors.
Certain types of assignments like multiple choice exercises, can be easily corrected automatically.
Other types of exercises, e.g. programming exercises, can hardly be assessed completely automatically.
Nevertheless, it is possible to support the assessment of such exercises by methods comparing
the handed-in solutions with exemplary solutions on different levels: structurally and functionally.

Flexible Knowledge Basis for Heterogeneous Data


With the Problem of handling big heterogeneous (Web based) data new approaches of storing, retrieving,
and reasoning were developed. These methods differ from the classical relational data model.
Methods from the semantic web field allow data bases which directly use xml structures for data representation or to use very flexible triple storages.