Data Science Poster & Pizza Session - Christmas Edition

16 Dec 2015

Students and postdocs carrying out research in data science will present posters about their current work, including drinks and pie2 (mince and pizza).

Date: Wednesday 16 December 2016

Time: 16:00-18:30

Location: Mini-Forum 1, Informatics Forum, University of Edinburgh


Xavier Cabezas, "Struggling with Traffic Lights: An MILP-Based Algorithm."

One of the most studied approaches to solve the synchronization of traffic lights on transport networks has been the maximization of the time during which a group of cars can start at one end of
a street and go to the other without stopping for a red light (bandwidth). One of the most influential approaches to tackle the problem is one by mixed integer linear programming (MILP) called MAXBAND. The high number of equality constraints and integers variables makes this problem a one challenging and this is why heuristic procedures have been the choice in recent work.
In this poster, we present a quick review of the model, bounds on the integers variables, and a easy way to formulate the loop constraints in the formulation. A MILP-TabuSearch-based Algorithm is also proposed and tested on grid graphs problems considering all the constraints detailed in MAXBAND. The computational evidence shows that the procedure yields a gain in time on relatively large instances when a VNS (variable neighborhood search) scheme is used.

Luca del Pero, "Articulated Motion Discovery using Pairs of Trajectories"

Konstantinos Georgatzis, "Discriminative Switching Linear Dynamical Systems applied to Physiological Condition Monitoring"

A Discriminative Switching Linear Dynamical System (DSLDS) applied to patient monitoring in Intensive Care Units (ICUs) is presented. The approach is based on identifying the state-of-health of a patient given their observed vital signs using a discriminative classifier, and then inferring their underlying physiological values conditioned on this status. The work builds on the Factorial Switching Linear Dynamical System (FSLDS) which has been previously used in a similar setting. The FSLDS is a generative model, whereas the DSLDS is a discriminative model. We demonstrate on two real-world datasets that the DSLDS is able to outperform the FSLDS in most cases of interest, and that an α-mixture of the two models achieves higher performance than either of the two models separately.

Michael Nicolson, "Clone sizes in growing populations: applications to metastasis formation"

Metastatic formation is modelled by a growing primary tumour population stochastically seeding metastatic clones, a so-called Luria-Delbrück model. The impact of the primary tumour growth on the metastatic clone sizes is investigated. Exponential, power-law and logistic growth are treated in detail. A "fat" tail is predicted in the metastatic size distribution and this is compared against empirical metastatic data.

George Papamakarios, "Distilling Model Knowledge"

Imagine we are given a knowledgeable model, such as a deep neural network, an ensemble of classifiers or a large collection of samples from a posterior distribution. Despite knowledgeable, the model may be cumbersome to work with; it may be large to store, expensive to evaluate or, if a generative model, intractable to do inference in. To address this problem, we present a framework for distilling the knowledge contained in the cumbersome model into a convenient model of our choosing that does the same job just as well. We show how our framework can be successfully used for compressing the size of models, for constructing compact predictors for Bayesian inference and for replacing intractable generative models by tractable ones.

M. Sam Ribeiro, "A multi-level representation of f0 using the continuous wavelet transform and the discrete cosine transform."

Data Science Pizza is a bi-monthly meeting for those carrying out research in data science to showcase their research and discuss ideas and socialise with researchers in other fields. The format will be a small number of posters plus food and drink. All those involved in data science research are encouraged to attend, from postgraduates to lecturing staff and research staff. Departments participating include: Informatics, Maths, and Engineering.