+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Donnerstag, 6. Oktober 2022, 15:00 Uhr
Ort: Raum 9222, Ahornstr. 55
Referent: Tim Hartmann M.Sc.
Lehrstuhl Informatik 1
Thema: Facility Location on Graphs
Abstract:
We study two closely related Facility Location problems on graphs where
all edges have unit length and where the facilities may also be
positioned in the interior of the edges. For delta-Dispersion the goal
is to position as many facilities as possible subject to the condition
that any two facilities have at least distance delta from each other.
For delta-Covering the goal is to cover the entire graph with the
minimum number of facilities; that is, we want to position as few
facilities as possible subject to the condition that every point on
every edge is at distance at most delta from one of these facilities.
We investigate the algorithmic complexity of these problems for every
real distance value delta. Further, we explore the complexity of these
problems with the solution size as parameter. Finally, we study
delta-Dispersion with parameters that measure the complexity of the
input graph, such as treewidth, treedepth and neighborhood diversity.
Es laden ein: die Dozentinnen und Dozenten der Informatik
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Freitag, 21. Oktober 2022, 14:30 Uhr
Ort: Raum 9U10 (2359|0.10), E3, Informatikzentrum, Ahornstr. 55
Referent: Rebecca Haehn M.Sc. (Theory of Hybrid Systems)
Thema: Optimisation and Analysis of Railway Timetables
under Consideration of Uncertainties
Abstract:
Railway systems are complex systems that are strongly affected by
uncertainties like weather, technical problems, or demand. Despite
these uncertainties, railway systems need to function efficiently. In
general this thesis aims to advance the consideration of uncertainty in
the railway planning process to optimally utilise the existing railway
network capacity. The focus in this thesis is on the delays that result
from the uncertain environmental conditions. To consider these in the
railway planning process, a symbolic simulation algorithm is proposed
to examine the delay propagation in a railway network for a given
timetable. This allows to estimate the timetable robustness and the
network capacity. Several performance indicators for railway timetables
that can be evaluated using the symbolic simulation are discussed. To
optimally utilise the network capacity also an algorithm to schedule
additional freight trains is presented.
The main contributions of this thesis are the following:
- An algorithm to schedule additional freight trains is presented,
to utilise the remaining network capacity without disturbing an
existing timetable.
- A novel symbolic simulation algorithm for railway timetables is
proposed. The algorithm receives as input a railway infrastructure
model, a corresponding timetable, and discrete primary delay dis-
tributions. It computes iteratively over time the delay propagation
in the given railway system. Symbolic expressions are used to
represent multiple possible values for the primary delays. This
enables to simulate all discrete primary delay combinations at once.
- An implementation of these algorithms is provided in C++ and
evaluated on some real-world railway infrastructure networks and
timetables based on the German railway system. The applicability
and functionality of the algorithms is demonstrated.
The proposed symbolic simulation algorithm is aimed to be a helpful
addition to existing railway timetable simulations, which are mostly
based on Monte Carlo simulation. In contrast to those, the symbolic
approach stores the history of specific train states, which can be used to
explain the occurring delays. In addition, the results of the symbolic
simulation are exact with respect to the input model and the discrete
primary delay distributions.
Es laden ein: die Dozentinnen und Dozenten der Informatik
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Dienstag, 4. October 2022, 10:00 Uhr
Ort: Raum 9222, Geb. E3, 2. Etage, Informatikzentrum, Ahornstr. 55
Referent: Shahid Khan M.Sc.
(Lehrstuhl Informatik 2)
Thema: Boolean-logic Driven Markov Processes - Explained. Analysed.
Verified.
Abstract:
Model-based dependability studies of engineering systems amount to
quantifying and improving dependability measures. These
measures include reliability, availability, maintainability and safety.
Two key ingredients of such studies are 1) models capturing the system
behaviour to an acceptable level of abstraction and 2) efficient and
accurate analysis techniques to quantify dependability measures.
Among the modelling techniques, static (or standard) fault trees
(SFTs) is a prominent dependability modelling language extensively used
by engineers to develop system models. However, it lacks expressive
power as it cannot model temporal dependencies of failures and has
limited support for repairs. Boolean logic-driven Markov processes
(BDMPs) and dynamic fault trees (DFTs) are two classical dynamic
extensions of SFTs that aim to mitigate the expressive power limitation
issue of SFTs. While DFTs are (generally) restricted to non-repairable
systems, BDMPs natively support repairs. The BDMP language has a
long-standing industrial usage history; the leading French electricity
utility company (EDF) extensively uses BDMPs to conduct dependability
studies.
Among the analysis techniques, probabilistic model checking is a
verification technique to determine the probability of a state-space
model satisfying or refuting a logical property. It combines efficient,
fully automated verification algorithms with numerical analysis. The
results of these automated verification procedures are numeric values
that dependability practitioners use to attain the reliability growth of
their systems.
The issue with BDMPs is that it lacks numerically exact analysis support and adequately
documented semantics. EDF provides two analysis tools for BDMPs: 1) a
Monte-Carlo simulator and 2) a sequence exploration
tool. Both tools provide approximate results, which may not be
acceptable for stringent dependability requirements of
safet-criticalsystems. The
semantics of BDMPs is essential for a comparative study with other
modelling languages, e.g., DFTs. This dissertation, developed in
collaboration with EDF, presents semantics and a model checker for BDMPs:
-- we propose Markov automaton- and generalised stochastic Petri
net-based semantics to BDMPs and empirically establish the accuracy,
-- we develop a probabilistic model checker for BDMPs and enhance the
scalability of the model checker using partial state-space exploration
techniques, and
-- we contrast BDMPs to DFTs and lift the repairable behaviour of BDMPs
to repairable DFTs.
This dissertation provides a holistic view of BDMPs. An exciting outcome
of this dissertation is that our model checker can analyse the Markovian
subset of the EDF-maintained Figaro language. This subset includes
capacity analysis diagrams, dynamic reliability block diagrams, electric
circuits, Petri nets, process diagrams, telecommunication networks, and
other EDF-developed Figaro-based modelling formalisms.
In this presentation, we provide: (1) a detailed introduction to
the BDMP formalism, (2) an insight of the BDMP model checker, and (3)
the semantics of BDMPs.
Es laden ein: die Dozentinnen und Dozenten der Informatik
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Dienstag, 20.09.2022, 10:00-11:00 Uhr
Der öffentliche Vortrag findet hybrid statt:
Raum: Raum 5053.2 (B-IT-Hörsaal)/Informatikzentrum, Ahornstraße 55
Zoom:
https://rwth.zoom.us/j/91824143381?pwd=dU5SNSt2TWtUMlZTOVFtRkczN1RvQT09
Meeting-ID: 918 2414 3381
Kenncode: 827047
Referent: Herr Thomas Osterland, M. Sc.
Lehrstuhl Informatik 5
Thema: Distributed Ledger Technology Processes
Abstract:
DLT offers new means to establish trust relationships within cooperative
processes among enterprises and thereby enables the implementation of use
cases with a lack of existing intermediaries. Prominent examples of use
cases are financial applications, provenance tracking and identity
management. All of them are contingent upon trust in business processes and
require the maintained security of the distributed ledger, that strongly
depends on the heterogeneity and robustness of the underlying ecosystem.
Participants need to trust the security of the ecosystem and the trust must
be maintained over time. Therefore, the distributed ledger technology is not
a simple component, that can be easily integrated into an existing
application unless the complex relationship between participants and their
trust into the underlying ecosystem is considered early in the engineering
process.
This thesis introduces a framework, that comprises a structured methodology
and tooling as a means to support conventional software processes that allow
software architects to address the specific requirements of distributed
ledger applications. We argue that these subjects are important to ensure
the quality and sustainability of an application, since 1) a distributed
ledger, that is not capable of handling future requirements of a use case
represents eventually a bottleneck and thus, harms the efficient and
reliable execution of business processes, 2) smart contracts with software
bugs, that either prevent the efficient execution of business processes or
favor one party over another will erode the trust that participants have
into the ecosystem, and 3) if the DLT is not capable of securing substantial
amounts of data use cases, that rely on the processing of data are
disadvantaged and the consequences for an ecosystem when handling large
amounts of data in the future are unforeseeable.
Es laden ein: die Dozentinnen und Dozenten der Informatik
_______________________________
Leany Maaßen
RWTH Aachen University
Lehrstuhl Informatik 5, LuFG Informatik 5
Prof. Dr. Stefan Decker, Prof. Dr. Matthias Jarke,
Prof. Gerhard Lakemeyer Ph.D.
Ahornstrasse 55
D-52074 Aachen
Tel: 0241-80-21509
Fax: 0241-80-22321
E-Mail: maassen(a)dbis.rwth-aachen.de
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Freitag, 26. August 2022, 13.00 Uhr
Der öffentliche Vortrag findet hybrid statt:
Ort: Raum 9222, Ahornstr. 55
Zoom:
https://rwth.zoom.us/j/92249056004?pwd=cHkwRW1lNTlmS1RXQW95Wmw1MjhFQT09
Referent: Janosch Fuchs M.Sc.
Lehrstuhl Informatik 1
Thema: Graph Exploration with Advice (and Online Crossing Minimization)
Abstract:
Online problems reveal their input instance piece by piece and require
an irrevocable decision each time a new piece is revealed. This scenario
models critical applications where a decision, once it is made, cannot
be changed afterwards and also influences future decisions. Like for
classical offline problems, it is common to make a worst-case analysis
to give a guarantee on the performance of an online algorithm. A
worst-case analysis is achieved by assuming that a malicious adversary,
knowing the behavior of the online algorithm, creates the input. Due to
the nature of uncertainty online algorithms rarely compute an optimal
solution. Thus, their performance is measured in terms of the
competitive ratio, which compares the solution computed by the online
algorithm to the optimal solution achievable when the whole instance is
known beforehand.
To measure the impact of the missing knowledge about the future the
classical online setting was extended with a helpful oracle providing
information about the future input. The online algorithm receives the
information by accessing a tape containing a binary bit string. The
number of bits that the algorithm reads during its computation is called
its advice complexity. Bounds on the advice complexity of an online
algorithm that optimally solves an online problem show how much
information about the future is needed to avoid any mistake. An obvious
upper bound for the advice complexity of all online problems is the size
of the binary encoding of the input instance or the optimal solution.
But most of the time it is possible to derive better bounds by encoding
only critical decisions which then reveals the core difficulty of an
online problem. Thus, the advice complexity strongly correlates to the
difficulty of an online problem and it can be used to measure the
difficulty of different online problems.
In the main part, we analyze the advice complexity of the graph
exploration problem. Here, an algorithm has to move an agent through a
graph from vertex to vertex using the edges such that the agent
traverses the cheapest or shortest tour that visits every vertex at
least once. The algorithm does not know the graph beforehand and each
time the agent is located at a new vertex the reachable neighborhood is
revealed together with the cost to reach them from the current location.
Already seen or visited vertices are recognizable. Due to the already
known tight upper bound of n\log(n) for the advice complexity of the
graph exploration problem on dense graphs, we focus on sparse graphs. We
show that a linear amount of advice is sufficient and also necessary to
solve the graph exploration problem on directed and undirected graphs.
Es laden ein: die Dozentinnen und Dozenten der Informatik
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Mittwoch, 10.08.2022, 14:00-15:00 Uhr
Ort: Raum 5053.2 (großer B-IT-Raum)/Informatikzentrum, Ahornstraße
55
Referent: Herr Zhijian Li , M.Sc.
Institute for Computational Genomics
Thema: Computational Method for Single-cell ATAC-seq Imputation and
Dimensionality Reduction
Abstract:
Single-cell sequencing assay for transposase-accessible chromatin
(scATAC-seq) allows mapping of regulatory variation of thousands of cells at
the single-cell resolution. However, analyzing the scATAC-seq data is
computationally challenging due to the sparsity, high dimensionality, and
the nature of the binary signal. In this thesis, we proposed scOpen, a
computational approach for quantification of single-cell open chromatin
status and reduction of the dimensionality based on the non-negative matrix
factorization (NMF) topic modeling. We demonstrated that scOpen can improve
several crucial downstream analysis steps of scATAC-seq, such as clustering,
visualization, cis-regulatory DNA interactions, and delineation of
regulatory features. We also applied scOpen to investigate the regulatory
programs that drive the development of chronic kidney disease (CKD).
Altogether, these results demonstrate that scOpen is a useful computational
approach in single-cell open chromatin data analysis.
Es laden ein: die Dozentinnen und Dozenten der Informatik
_______________________________
Leany Maaßen
RWTH Aachen University
Lehrstuhl Informatik 5, LuFG Informatik 5
Prof. Dr. Stefan Decker, Prof. Dr. Matthias Jarke,
Prof. Gerhard Lakemeyer Ph.D.
Ahornstrasse 55
D-52074 Aachen
Tel: 0241-80-21509
Fax: 0241-80-22321
E-Mail: maassen(a)dbis.rwth-aachen.de
Dear all,
this is a reminder for Gerhard Lakemeyer's talk on Tractable Reasoning
in First-Order Knowledge Bases
<https://www.unravel.rwth-aachen.de/cms/UnRAVeL/Das-Graduiertenkolleg/Aktuel…>
taking place *today at 16:30* in room 5053.2 and on Zoom. Please find
the details below.
> In knowledge representation, obtaining a notion of belief which is
> tractable, expressive, and eventually complete has been a somewhat
> elusive goal. Expressivity here means that an agent should be able to
> hold arbitrary beliefs in a very expressive language like that of
> first-order logic, but without being required to perform full logical
> reasoning on those beliefs. Eventual completeness means that any
> logical consequence of what is believed will eventually come to be
> believed, given enough reasoning effort.
>
> Tractability in a first-order setting has been a research topic for
> many years, but in most cases limitations were needed on the form of
> what was believed, and eventual completeness was so far restricted to
> the propositional case. In this talk I present a novel logic of
> limited belief, which has all three desired properties.
Part of the programme of the research training group UnRAVeL is a series
of introductory lectures on the topics of "randomness" and "uncertainty"
in UnRAVeL’s research thrusts: Algorithms and complexity, verification,
logic and languages, and their application scenarios. The main aim is to
provide doctoral researchers as well as master students a broad overview
of the subjects of UnRAVeL.
Science undergoes continuous change and lives from the constant quest
for novel and better results, which are presented at conferences and in
journals. This year, 10 UnRAVeL professors will present some of their
most recent research successes.
Everyone interested, in particular doctoral researchers and master
students, are invited to attend the UnRAVeL lecture series 2022 and
engage in discussions with the researchers.
The talks take place on Tuesdays, 16:30–18:00 in room 5053.2 in the
ground floor of building E2. All events are hybrid. To join remotely,
please use
https://rwth.zoom.us/j/96003885007?pwd=aUczMVdVU0ZXVGtQUFpwQnJHQUFhUT09
/ Meeting ID: 960 0388 5007 / Passcode: 273710
We are looking forward to seeing many of you in the UnRAVeL survey
lecture "What's New in UnRAVeL?".
Best regards,
Andreas Klinger, Birgit Willms, and Tim Seppelt
Logo
Dear all,
this is a reminder for Britta Peis' talk on Stackelberg Network Pricing
Games
<https://www.unravel.rwth-aachen.de/go/id/tbciy?lidx=1#aaaaaaaaaatbcjr>
taking place *today at 16:30* in room 5053.2 and on Zoom. Please find
the details below.
> We study a variant of bilevel games, termed Stackelberg network
> pricing games, in which one distinguished player, the leader, can set
> prices on a subset of the edges or vertices in the underlying network.
> The remaining edges or vertices are assigned a fixed costs. Based on
> the leader’s decision, one or several followers optimize a
> polynomial-time solvable combinatorial optimization problem. That is,
> after the leader decided on the prices, each follower individually
> selects an optimal solution (e.g. a shortest path, a max weight
> matching, or a min cost spanning tree) based on the fixed costs and
> the leader’s prices. The leader’s goal is to select the prices such
> that the resulting revenue, which is determined by the sold items and
> their prices, is maximised.
>
> Briest et al. and Balcan et al. independently showed that the maximum
> revenue can be approximated to a factor of (roughly) log k, where k is
> the number of priceable elements. In the lecture, we discuss
> algorithms and complexity results for Stackelberg network pricing
> games in general, and Stackelberg Max Closure and Stackelberg
> Bipartite Min Weight Vertex Cover, in particular.
>
> Parts of the talk are based on joint work with Karsten Jungnitsch and
> Marc Schröder
Part of the programme of the research training group UnRAVeL is a series
of introductory lectures on the topics of "randomness" and "uncertainty"
in UnRAVeL’s research thrusts: Algorithms and complexity, verification,
logic and languages, and their application scenarios. The main aim is to
provide doctoral researchers as well as master students a broad overview
of the subjects of UnRAVeL.
Science undergoes continuous change and lives from the constant quest
for novel and better results, which are presented at conferences and in
journals. This year, 10 UnRAVeL professors will present some of their
most recent research successes.
Everyone interested, in particular doctoral researchers and master
students, are invited to attend the UnRAVeL lecture series 2022 and
engage in discussions with the researchers.
The talks take place on Tuesdays, 16:30–18:00 in room 5053.2 in the
ground floor of building E2. All events are hybrid. To join remotely,
please use
https://rwth.zoom.us/j/96003885007?pwd=aUczMVdVU0ZXVGtQUFpwQnJHQUFhUT09
/ Meeting ID: 960 0388 5007 / Passcode: 273710
Please find a list of all upcoming talks on the UnRAVeL website
<https://www.unravel.rwth-aachen.de/cms/UnRAVeL/Studium/~pzix/Ringvorlesung-…>
and below:
* 28/06/2022 Britta Peis: Stackelberg Network Pricing Games
* 05/07/2022 Gerhard Lakemeyer: Tractable Reasoning in First-Order
Knowledge Bases
We are looking forward to seeing many of you in the UnRAVeL survey
lecture "What's New in UnRAVeL?".
Best regards,
Andreas Klinger, Birgit Willms, and Tim Seppelt
Logo
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Dienstag, 12. Juli 2022, 9.00 Uhr
Zoom: https://rwth.zoom.us/j/99072269440?pwd=bXNrOXlwSUhtd3IvVVlhNytURUY2UT09
Referent: Arvid Butting M.Sc.
Lehrstuhl Informatik 3
Thema: Systematic Composition of Language Components in MontiCore
Abstract:
In model-driven development (MDD), models are central software engineering artifacts. MDD is applied to various domains such as avionics, law, mechanical engineering, or robotics, in which the domain engineers are not always software engineers. To this end, modelers should specify models in a notation close to the application domain, which is achieved by employing domain-specific modeling languages (DSMLs). In complex modern software applications, different aspects of an application are modeled with numerous integrated models. The models conform to heterogeneous, integrated DSMLs that can assure consistency between the models of an application.
Ad-hoc development of DSMLs is a time-consuming and error-prone process. Systematic and "off-the-shelf" black-box reuse of DSMLs or parts of it supports engineering DSMLs faster and more reliably. In black-box reuse, unlike reuse via clone-and-own, the reused parts remain unchanged and do not result in co-existing clones. Such reuse requires language engineers to be able to integrate DSMLs through different forms of language composition. Current approaches for engineering DSMLs often rely on generic language infrastructure, which complicates compatibility checks between the infrastructures of languages that are to be composed. Approaches for modularization of DSMLs typically focus on the conceptual parts of a language rather than on their realizations.
This talk describes an approach for realizing modular language components that can be composed via their symbol tables to realize language product lines with the language workbench MontiCore. The proposed language components identify the entirety of source code artifacts that realize a DSML. The DSMLs rely on kind-typed symbol tables that assure language compatibility during language composition. Language composition via symbol tables is lightweight because language infrastructures are only loosely coupled. An approach for persisting symbol tables further decouples language infrastructures from another and increases the performance for type and consistency checking between models that conform to different DSMLs. With the approach for language product lines, language components can be composed systematically and undesired compositions can be avoided. Typed and persisted symbol tables, language components, and language product lines as presented in this talk aim to realize DSML engineering in the large.
Es laden ein: die Dozentinnen und Dozenten der Informatik
Sehr geehrte Damen und Herren,
leider hat sich ein Fehler im Zoom-Link eingeschlichen. Bitte beachten Sie
deshalb die nachfolgenden Informationen sowie den entsprechenden Link.
Vielen Dank!
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Dienstag, 28.06.2022, 10:00-11:00 Uhr
Der öffentliche Vortrag findet hybrid statt:
Raum: Raum 5053.1 (kleiner B-IT-Raum)/Informatikzentrum, Ahornstraße 55
Zoom: https://zoom.us/j/91338931025?pwd=SDN3TWh2c0ZTb2xjWGkwdHZYLzQ5QT09
Referent: Herr Md Rezaul Karim, Master of Engineering
Lehrstuhl Informatik 5
Thema: Interpreting Black-Box Machine Learning Models with Decision
Rules and Knowledge Graph Reasoning
Abstract:
Machine learning (ML) algorithms are increasingly used to solve complex
problems. However, due to high non-linear and higher-order interactions
between features, complex ML models become black-box methods - which means
it is not known how certain predictions are made. This may not be
acceptable in many situations (e.g., in clinical situations where AI may
significantly impact human lives). With the EU GDPR explainability has not
only become a desirable property of AI but also a legal requirement. An
interpretable ML model can outline how input instances are mapped into
certain outputs by identifying statistically significant features.
Literature pointed out that complex ML models tend to be less interpretable,
showing a trade-off between accuracy and interpretability. This thesis aims
to improve the interpretability and explainability of black-box ML models
without sacrificing significant predictive accuracy. As a starting point,
using a black-box multimodal neural network, representation learning is
performed on multimodal data in order to use the learned representation for
the classification task. To improve the interpretability of the learned
black-box model, different interpretable ML methods such as probing,
perturbing, and model surrogation techniques are applied. An interpretable
surrogate model is trained to approximate the behavior of the back-box
model. The surrogate model is used to generate explanations in terms of
decision rules and counterfactuals. To add symbolic reasoning capability to
the black-box model, a domain-specific knowledge graph (KG) is constructed
by integrating knowledge and facts from scientific literature. A semantic
reasoner is then used to validate the association of significant features
with different classes based on relations it learned from the KG.
Evidence-based decision rules are generated by combining the reasoning with
the predictions from the black-box model. The quantitative evaluation shows
that the proposed approach achieves an average accuracy of 96.25% on the
test dataset. It can also provide human-interpretable explanations of the
decisions in the form of counterfactual rules and evidence-based decision
rules. The quality of the explanations is evaluated in terms of
comprehensiveness and sufficiency.
Es laden ein: die Dozentinnen und Dozenten der Informatik
--
Romina Reddig
_______________________________
Romina Reddig
RWTH Aachen University
Lehrstuhl Informatik 5, LuFG Informatik 5
Prof. Dr. Stefan Decker, Prof. Dr. Matthias Jarke,
Prof. Gerhard Lakemeyer Ph.D.
Ahornstrasse 55
D-52074 Aachen
Tel: 0241-80-21501
Fax: 0241-80-22321
E-Mail: reddig(a)dbis.rwth-aachen.de <mailto:reddig@dbis.rwth-aachen.de>
Von: Romina Reddig <reddig(a)dbis.rwth-aachen.de>
Gesendet: Montag, 20. Juni 2022 11:41
An: assistenten(a)informatik.rwth-aachen.de;
infprof(a)informatik.rwth-aachen.de; vortraege(a)informatik.rwth-aachen.de;
sekretariate(a)informatik.rwth-aachen.de; webteam(a)informatik.rwth-aachen.de;
biblio(a)informatik.rwth-aachen.de
Betreff: Einladung Promotionsverteidigung Rezaul Karim, 28. Juni 2022, 10:00
Uhr
+**********************************************************************
*
*
* Einladung
*
*
*
* Informatik-Oberseminar
*
*
*
+**********************************************************************
Zeit: Dienstag, 28.06.2022, 10:00-11:00 Uhr
Der öffentliche Vortrag findet hybrid statt:
Raum: Raum 5053.1 (kleiner B-IT-Raum)/Informatikzentrum, Ahornstraße 55
Zoom: https://zoom.us/j/94726175294?pwd=QXJVeGVvZ09wV000OU53QkdrU0RRdz09
Referent: Herr Md Rezaul Karim, Master of Engineering
Lehrstuhl Informatik 5
Thema: Interpreting Black-Box Machine Learning Models with Decision
Rules and Knowledge Graph Reasoning
Abstract:
Machine learning (ML) algorithms are increasingly used to solve complex
problems. However, due to high non-linear and higher-order interactions
between features, complex ML models become black-box methods - which means
it is not known how certain predictions are made. This may not be
acceptable in many situations (e.g., in clinical situations where AI may
significantly impact human lives). With the EU GDPR explainability has not
only become a desirable property of AI but also a legal requirement. An
interpretable ML model can outline how input instances are mapped into
certain outputs by identifying statistically significant features.
Literature pointed out that complex ML models tend to be less interpretable,
showing a trade-off between accuracy and interpretability. This thesis aims
to improve the interpretability and explainability of black-box ML models
without sacrificing significant predictive accuracy. As a starting point,
using a black-box multimodal neural network, representation learning is
performed on multimodal data in order to use the learned representation for
the classification task. To improve the interpretability of the learned
black-box model, different interpretable ML methods such as probing,
perturbing, and model surrogation techniques are applied. An interpretable
surrogate model is trained to approximate the behavior of the back-box
model. The surrogate model is used to generate explanations in terms of
decision rules and counterfactuals. To add symbolic reasoning capability to
the black-box model, a domain-specific knowledge graph (KG) is constructed
by integrating knowledge and facts from scientific literature. A semantic
reasoner is then used to validate the association of significant features
with different classes based on relations it learned from the KG.
Evidence-based decision rules are generated by combining the reasoning with
the predictions from the black-box model. The quantitative evaluation shows
that the proposed approach achieves an average accuracy of 96.25% on the
test dataset. It can also provide human-interpretable explanations of the
decisions in the form of counterfactual rules and evidence-based decision
rules. The quality of the explanations is evaluated in terms of
comprehensiveness and sufficiency.
Es laden ein: die Dozentinnen und Dozenten der Informatik
--
Romina Reddig
_______________________________
Romina Reddig
RWTH Aachen University
Lehrstuhl Informatik 5, LuFG Informatik 5
Prof. Dr. Stefan Decker, Prof. Dr. Matthias Jarke,
Prof. Gerhard Lakemeyer Ph.D.
Ahornstrasse 55
D-52074 Aachen
Tel: 0241-80-21501
Fax: 0241-80-22321
E-Mail: reddig(a)dbis.rwth-aachen.de <mailto:reddig@dbis.rwth-aachen.de>