Best PhD thesis, best paper and best poster prizes 2016Published: Monday, 21 November 2016
Winners from the School of Computer Science research symposium 2016
The 2016 Computer Science research symposium was held at the University of Manchester at the beginning of November with support from IBM. Over 80 PhD students from the School of Computer Science presented their research in a wide range of fields including image processing, networks, data management, unstructured data, security, software engineering, optimisation, learning from data, multicore, chip design and hardware. Congratulations to the winners of the best PhD thesis, paper and poster prizes as follows:
- Patrick Koopmann and Kostas Sechidis: Joint winners of the best PhD thesis prize
- Nikos Nikolaou, best paper award (picture above)
- James Knight, best paper runner-up
- Sukru Eraslan, best paper runner-up
- Robert James, best poster
- Ainur Begalinova, best poster runner-up
Best thesis: Patrick Koopmann
The external examiner commented on Patrick Koopmann's PhD thesis, Practical Uniform Interpolation for Expressive Description Logics:
“Ontologies are a means to enrich computer science applications with domain knowledge, which is particularly relevant in areas such as the Semantic Web, Data Science, and Data Integration. They have received significant and long- lasting interest from the CS research community. In his thesis, Patrick studies the computation of uniform interpolants of ontologies that are formulated in a description logic (DL), that is, given a DL ontology, one wants to eliminate some of the symbols contained in it without altering the meaning of the remaining symbols. For example, one might want to trim down a thematically broad medical ontology by eliminating all symbols except those that describe terms from anatomy. Uniform interpolation is a foundational reasoning service that has applications in many areas including ontology reuse, ontology comprehension, and information hiding.
The aim of Patrick's thesis is to enable the efficient computation of uniform interpolants in real-world applications. This is a challenging endeavour since the addressed problem is computationally very hard [...] and, in general, uniform interpolants are not even guaranteed to exist. In fact, no practically feasible procedures capable of computing uniform interpolants for expressive DL ontologies were known before the work of Patrick since all existing algorithms involved some brute-force enumerations.
The thesis provides a rich source of techniques and (counter)examples for the computation of uniform interpolants with fixpoints. Patrick supports the presentation of his calculi with a large number of well-chosen examples, thus making the thesis highly accessible despite significant technical intricacies. The theoretical part of the thesis alone certainly goes beyond what one would expect of a typical PhD thesis. The additional and very substantial implementation and experiments part completes and rounds off the thesis. In fact, the comprehensive experimental evaluation impressively demonstrates that the calculi are indeed practically feasible, and thus the thesis fully accomplishes its goal.
...Because of the many interesting applications of this reasoning service ... the thesis will generate a lot of interest and receive significant uptake.”
Best thesis: Kostas Sechidis
The external examiner commented on Kostas Sechidis’s PhD thesis, Hypothesis Testing and Feature Selection in Semi-Supervised Data:
“The thesis addresses a very important problem of selecting “informative” features among a (possibly very) large set of available features in the classification setting. Unlike in the classical classification framework, where each data item in the training set is expected to be labeled as belonging to one of the classes under consideration, Kostas chose to address a much more difficult situation of partially labeled data. In particular, two scenarios of partially labeled data were considered: (1) semi-supervised setting, where examples with labels and without labels can be found in all (in this case two) classes; and (2) positive-unlabeled data setting, where the labeled set contains only examples from one (e.g. positive) class.
The importance of this work cannot be overestimated, since in many real world problems collecting unlabeled examples is much more straightforward than collecting the corresponding labels. This can be because the labels are simply not known for most data, or are difficult/expensive to obtain. In such situations, having provably correct and principled methods for feature selection is imperative.
Kostas’ work represents a major, serious and rigorous effort along the lines of feature selection in partially labeled data in the frameworks of statistical testing and information theory. I was very impressed by the depth and variety of the obtained results. In addition, the thesis is highly readable and very well presented and organized.
The work represents a major step forward. ...the work makes an outstanding contribution to scientific knowledge with deep theoretical results well supported by empirical experiments.”
Best Paper Award: Winner: Nikos Nikolaou
Nikos won this award for his paper “Cost Sensitive Boosting Algorithms: Do We really need them?” Nikos Nikolaou, Nara Edakunni Meelis Kull, Peter Flach, and Gavin Brown, Machine Learning Journal, 104(2), 359 – 384, 2016.
The paper unifies two decades of research on a very popular class of Machine Learning algorithms called “boosting”, in particular looking at cost-sensitive versions. The key finding is that an entire class of techniques - 15 algorithmic variants - studied by en entire community over the past 20 years, can effectively be discarded if the original algorithm is correctly calibrated. The work was selected for plenary presentation at ECML 2016, indicating it stands in the top 2% of submitted papers this year. One of the reviewers said of it
“The paper is well written and a pleasure to read. The analysis is insightful … and will serve as an excellent reference for future research in this area."
Best Paper Runner-up: James C. Knight
James won this award for his paper Large-Scale Simulations of Plastic Neural Networks on Neuromorphic Hardware, James Courtney Knight, Philip Joseph Tully, Bernhard A. Kaplan, Anders Lansner, Steve B. Furber, Frontiers in Neuroanatomy, 10(37), 2016.
This paper demonstrates a large scale implementation of plasticity in neural circuits on SpiNNaker. The energy efficiency is demonstrated to be 45 times more efficient than a Cray supercomputer implementation. The reviewers said of it:
“This is a well executed study, … well executed, constructed and presented .. and extremely valuable for the neuroscience community"
The paper was selected by the Faculty of Science and Engineering as an example of a “World Leading Paper”, and featured on the faculty website, saying that the work “will provide a valuable resource to aid future developments the fields of neuroscience, robotics and computer science.
Best Paper Runner-up: Sukru Eraslan
Sukru won this award for his paper “Scanpath Trend Analysis on Web Pages:Clustering Eye Tracking Scanpaths”, Yeliz Yesilada, Simon Harper, Transactions on the Web, in press. The paper introduces the Scanpath Trend Analysis (STA) technique, which constructs a "trending" scanpath of the most visited page elements taking into account all of the visual elements visited by all users, in any order. One reviewer said of it
"This work is very interesting and well motivated. The work will certainly inspire further research, especially in web design (front & back end), user studies involving eye tracking and related eye tracking data analysis."
Robert James was awarded the best poster while Ainur Begalinova was the best poster runner-up.