Makerere University School of Computing and I.T.
Artificial Intelligence in the Developing World
Past Seminars

First Makerere Workshop on Social Systems and Computation, 3-10 December, 2010.

Back to current seminars

26 February 2015

Speaker: John Quinn
Speech Recognition and Application to Ugandan English
Automatic speech recognition (ASR) is a mature field for the world's most widely spoken languages. However, for less common languages, such as those spoken in Uganda, there is little or no available speech technology. This talk will give an overview of the methods and types of dynamical models used for ASR, and show some work in progress on applying these locally. This is beginning with Ugandan pronunciation of English, later to be extended to Luganda and Acholi.

19 February 2015

Speaker: Rose Nakibuule
Effect of size of region of interest (ROI) on accuracy average speed estimation
Traffic average Speed is a critical parameter in traffic flow monitoring systems as it forms a basis for traffic flow management on roadways, monitoring/prevention of accidents and providing secure transportation. Accurate estimation of traffic flow average speed help in making correct decisions on travel times, precess of accidents and traffic law violating vehicles. In this talk, we investigate the effect of size of region of interest (ROI) that is size of road segment being monitored on the accuracy of average speed estimation.

12 February 2015

Speaker: David Byansi
Comparative Study of White Blood Cell Detection and Classification with Computer Vision Techniques Using Different Imaging Devices

Microscopes using blood smear are the most widely used methods of white blood cell counting, which sometimes results in error, they are time consuming, labour intensive and expensive and the 100-cell differential count is often criticized for its statistical shortcomings because of its small sampling size.
A number of computer vision techniques have been applied to the process of automatic detection and classification of these white blood cells image taken from microscopic stained blood smear using cameras. However no literature has talked of mobile phone cameras being used for image collection. Yet in developing countries especially Uganda, using dedicated cameras for microscopes are expensive and mobile phone camera are predominantly cheap.
Therefore this research seeks to compare images collected using different imaging devices such a motic camera and a mobile phone (ZTE phone) camera in order to find out whether the accuracy of results obtained from microscopic blood smear images collected using a mobile phone camera can be close to that of the motic camera for image processing in white blood cell detection and classification.
This is done by using SIFT and Surf feature detection to compare whether the images from both cameras are the same. Then Annotation done using the software developed by PASCAL Visual Object Classes challenge. Feature detection and extraction done for both sets of images by determine the center of white blood cell with the help of the bounding boxes created by software and compare with each image patch to see whether the center of the white blood cell is within that patch.

Classification of each class will be done and accuracy of the classifier analyzed using the precision recall curve. The comparison of the sets of images from the two cameras will be done by looking at the results of curves of both classes of each classifier since a number of classifier will be used and the results of the best classifier will give us the conclusion whether the mobile phone camera can be used for image collection.

05 February 2015

Presenter: Deborah Mudali

29 January 2015

Speaker: Solomon Nsumba

22 January 2015

No Meeting -- Graduation

08 January 2015

Presenter: Fred Kiwanuka

18 December 2014

Speaker: M-Crops team
Update on the data collection by the M-Crops team
As we come to the end of year we can talk about:-
* progress made by AI Lab in 2014,
* targets/plans for the coming year 2015,
* AI Lab organisation
* get together
* and any other issue members think is important to the wellbeing and growth of AI research group.

04 December 2014

Speaker: John Quinn

20 November 2014

Speakers: Martin Mubangizi and Rose Nakibuule

13 November 2014

Speaker: Jovia Tuhaise and Godliver Owomugisha

23 October 2014

Speaker: Lukyamuzi Andrew

16 October 2014

Speaker: Ricardo Andrade-Pacheco

02 October 2014

Speaker: Micheal Smith

25 September 2014

Speaker: Julianne Sansa Otim

18 September 2014

Speaker: Martin Mubangizi

08 September 2014

Speaker: Jonathan Ledgard (Director of Future Africa at the Swiss Federal Institute of Technology in Lausanne (EPFL) and Africa correspondent at large of The Economist.)
Topic: Cargo drones, the future, and Uganda.

04 September 2014

Speaker:Dr. Feese
Activity recognition for monitoring group activity

28 August 2014

Speaker: John Quinn

21 August 2014

Speaker: Yonasi Safari

14 August 2014

Speaker: Owomugisha Godliver

07 August 2014

Speaker: David Byansi

31 July 2014

Speaker: Rahman Sanya

24 July 2014

Speaker: Deborah Mudali
Title: Decision Tree Classification of Parkinsonian Syndromes
Early diagnosis and differentiation of Parkinsonian syndromes is hard. In the hope to classify subject brain images,
features are derived and used to train decision tree classifiers. These classifiers are then validated to determine
how they will perform given features from subject images in the test set.

17 July 2014

Speaker: Brian Muchake
Topic: Cross Language information extraction between Lusoga and English language

10 July 2014

Speaker: Richard Ssekibuule

03 July 2014

Speaker: Richard Ssekibuule

26 June 2014

Speaker: Chris Williams (University of Edinburgh)

19 June 2014

Speaker: Margret Kayendo

12 June 2014

Speaker: Jay Taneja, IBM Research Nairobi
A Tool for Evaluating Options to Unlock Off-Grid, Distributed Electrification in Africa

05 June 2014

Speaker: Mike Smith

29 May 2014

Speaker: Martin Mubangizi

22 May 2014

Speaker: Martin Mubangizi

15 May 2014

Speaker: Gregg Zachary (Arizona State University)

08 May 2014

Speaker: to be announced

01 May 2014

Speaker: To be announced

24 April 2014

Speaker: John Quinn

17 April 2014

Speaker: Andrew Lukyamuzi

10 April 2014

Speaker: Jacob Katende

03 April 2014

Speaker: Rahman Sanya
Computational Analysis, Prediction, and Visualization of Mycobacterium Tuberculosis Complex (MTBC) Genotype Epidemiology Data from Uganda

27 March 2014

Speaker: Peter Nabende

20 March 2014

Speaker: Rose Nakibuule

13 March 2014

Speaker: Ernest Mwebaze

06 March 2014

Speaker: Anthony Gidudu
Image Classification in Remote Sensing Imagery

27 February 2014

Informal discussion on famine prediction

13 February 2014

Speaker: Wanjiku Nganga (University of Nairobi/Resilient Africa Network, Makerere)
Interlingua-based Swahili Machine Translation

15 January 2014

Speaker: Richard Ssekibuule
Review of relevant papers presented at the ACM DEV meeting

08 January 2014

Speaker: John Quinn

19 December 2013

Speaker: Joviah Tuhaise

12 December 2013

Speaker: Patrick Doherty from Linkoping University Sweden

05 December 2013

Speaker: Jennifer Aduwo

29 November 2013

Speaker: Martin Mubangizi

21 November 2013

Speaker: Peter Nabende

14 November 2013

Speaker: Richard Sekibuule

07 November 2013

Speaker: Martin Mubangizi

31 October 2013

Speaker: Albert Bitwire

24 October 2013

Speaker: Ernest Mwebaze

17 October 2013

Speaker: Peter Nabende

10 October 2013

Speaker: John Quinn

03 October 2013

Speaker: Visitor from University of Stellenbosch

26 September 2013

Speaker: Godliver Owomugisha

19 September 2013

Presenter: Martin Mubangizi

12 September 2013

Presenter: Martin Mubangizi

04 July 2013

Speaker: Martin Mubangizi Gordon

27 June 2013

Speaker: Ernest Mwebaze

20 June 2013

Speaker: Richard Ssekibuule

13 June 2013

Speaker: Maxwell Omwenga

06 June 2013

Video Seminar

30 May 2013

Speaker: Aloysius Ochola
Automated Classification of Short Messages Service (SMS)

23 May 2013

Speaker: John Quinn

16 May 2013

Speaker: Rose Nakibuule

09 May 2013

Speaker: Martin Mubangizi

02 May 2013

Presenter: Peter Nabende

25 April 2013

Presenter: Anthony Gidudu
A presenation on Remote Sensing

18 April 2013

Speaker: BSE13-12
Software egineering group doing a project in crop disease survillance system.
We shall use this time to also discuss about the video competition.

11 April 2013

Speaker: Martin Mubangizi

04 April 2013

Speaker: William Senfuma

28 March 2013

Video seminar: Toby Dylan Hocking, Tokyo Institute of Technology
Learning penalties for change-point detection using max-margin interval regression

21 March 2013

Speaker: John Wekesa

14 March 2013

Speaker: John Quinn

07 March 2013

Speaker: Mark Musumba

28 February 2013

Moderator: Martin Mubangizi
Brain storming session on AAAI Video Competitions

21 February 2013

Speaker: Joseph Lwomwa

14 February 2013

Speaker: Mary Nsabagwa

07 February 2013

Speaker: Washington Okori

31 January 2013

Speaker: Jovia Tuhaise

24 January 2013

Speaker: Albert Bitwire

23 January 2013

Speaker: Miguel Luengo Oroz - from UN Global Pluse

17 January 2013

Speaker: Martin Mubangizi

16 January 2013

Speaker: Richard Sekibuule

10 January 2013

Speaker: Godliver Owomugisha

09 January 2013

Speaker: John Quinn

03 January 2013

Speaker: Rose Nakibuule

20 December 2012

Speaker: Jennifer Aduwo

22 November 2012

Presenter: Prof Michael Vaganov
Application of AI in Game Programming

15 November 2012

Speaker: Martin Mubangizi

25 October 2012

Speaker to be confirmed

18 October 2012

Speaker : Fred Kiwanuka

11 October 2012

Speaker: Erik Urbach

04 October 2012

Speaker: John Quinn

27 September 2012

Speakers: Reinier Battenberg and Ketty Adoch (Mountbatten)

20 September 2012

Presenter: Kenneth Bwire

13 September 2012

Speaker: Martin Mubangizi
A Dive into Mapping: The world takes on a shape of an imperfect sphere (an ellipsoid with a 3D surface), but it is always presented on a 2D planar surface (paper maps or digital GIS) . This mapping from a 3D surface to a 2D surface is known as projection. These projections introduce some distortions in the original properties of the 3D surface. These properties include area, shape, distance and direction. A number of map projections have been proposed to conserve particular properties. I will be discussing two projections: Universal Transverse Mercator (UTM) and Albers Equal Area Conic.

06 September 2012

Presenter: Allan Tucker (Brunel University, London)

30 August 2012

Presenter: Rose Nakibuule

23 August 2012

Presenter: Ernest Mwebaze

13 August 2012

Speaker: Mobile Monday Kampala meeting on spatial data processing

09 August 2012

Presenter: Martin Mubangizi
AAAI 2012 highlights

05 August 2012

Presenter: Florence Tushabe

26 July 2012

Presenter: Charles Fox, University of Sheffield
Origin-Destination Analysis on the London Orbital Automated Number Plate Recognition Network

19 July 2012

Presenter: Richard Ssekibuule

28 June 2012

Presenter: Peter Nabende

21 June 2012

Presenter: John Quinn

14 June 2012

Presenter: Richard Ssekibuule

14 June 2012

Presenter: Richard Ssekibuule

07 June 2012

Presenter: Ernest Mwebaze

31 May 2012

Presenter: Rose Nakibuule

29 May 2012

Presenter: Jen Ziemke, Crisis Mappers Network
Informal introduction to crisis mapping

24 May 2012

Presenter: Martin Mubangizi

17 May 2012

Presenter: Mary Nsabagwa
Embedded systems in traffic monitoring

10 May 2012

Presenter: Ibrahim Bbossa
Computer vision on smartphones, with an example of handwritten character detection

03 May 2012

Presenter: John Quinn
Meta research group meeting

26 April 2012

Presenter: Ronald Wesonga
Parameterized framework for the analysis of probabilities of aircraft delay at an airport

19 April 2012

Presenter: Ronald Wesonga
Parameterized framework for the analysis of probabilities of aircraft delay at an airport

05 April 2012

Presenter: Florence Tushabe

29 March 2012

Presenter: Peter Nabende

22 March 2012

Presenter: Rose Nakibuule
Urban Traffic Management Systems

15 March 2012

Presenter: Richard Ssekibuule
The intersection between security and machine learning

01 March 2012

Presenter: Martin Mubangizi

23 February 2012

Presenter: Ernest Mwebaze

16 February 2012

Presenter: John Quinn
We'll look through the following paper:
Road Traffic Congestion in the Developing World. Jain et al, ACM DEV 2012.

27 January 2012

Presenter: Mark Musumba
This lecture will be in two parts. The first section will be an
introduction to Texas A&M University and the opportunities that are
available to prospective students. This portion will provide more
detailed information about the research and funding opportunities that
students can explore at the graduate level. In the second section of
the lecture, I will talk about my academic journey and motivation to
pursue a PhD. I will talk about my current research work and journal
publications. This section will be concluded with a presentation of my
current research paper on price dynamics; “Price Transmission to
Ugandan Coffee Growers in a Liberalized Market.” that is under review
at the Development Policy Review Journal (if time permits).

19 January 2012

Presenter: Athina Spiliopoulou
Title: The Restricted Boltzmann Machine and Deep Learning: Overview
and Applications
Abstract: Recently, many models based on the Restricted Boltzmann
Machine (RBM) have been proposed and successfully applied to a large
variety of learning problems. The RBM is a type of Neural Network that
has restricted connectivity and specific properties. One of its main
advantages is that we can create models with deep architecture by
consecutively stacking a new RBM on top of an RBM that we have already
trained. This allows us to learn deep models in a greedy, layer-wise
manner, and thus exploit the representational power of deep
architectures without significantly increasing the computational
complexity of our method.
In this talk I will first describe the Restricted Boltzmann Machine
and the way we can use it to create deep networks. Then I will discuss
successful applications of this framework from the literature and
finally I will describe how I have used the RBM to learn a generative
model for melodic sequences.

08 December 2011

Presenter: Ernest Mwebaze

01 December 2011

Presenter: Rose Nakibuule

24 November 2011

Presenter: Peter Nabende
Mining transliterations from Wikipedia using dynamic Bayesian networks

17 November 2011

Presenter: Martin Mubangizi

10 November 2011

Presenter: Frederick Ssemakula
Human intrusion detection with computer vision

03 November 2011

Presenter: John Quinn
Reading session: Green Driver: AI in a Microcosm. Jim Apple, Paul Chang, Aran Clauson, Heidi Dixon, Hiba Fakhoury, Matt Ginsberg, Erin Keenan, Alex Leighton, Kevin Scavezze and Bryan Smith, AAAI 2011.

27 October 2011

Presenter: George Okori
Prior knowledge contribution in probabilistic prediction of famine

20 October 2011

Presenter: Richard Ssekibuule
Auction and reputation system design for African commodity trading

13 October 2011

Presenter: Ernest Mwebaze

06 October 2011

Presenter: Catherine Ikae
Automatic Diagnosis of Malaria with Computer Vision

29 September 2011

Presenter: John Bosco Asiimwe
Disease mapping techniques for application to under-five mortality data in Uganda (Part 2 of talk earlier this year)

04 August 2011

Presenter: John Quinn
Automatically learning features for object detection

28 July 2011

Group discussion: Bayesian inference session 2

21 July 2011

Presenter: Martin Mubangizi

14 July 2011

Presenter: Rose Nakibuule

07 July 2011

Bayesian inference session

23 June 2011

Presenter: John Bosco Asiimwe
Small area estimation techniques for application to under-five mortality (disease mapping) data in Uganda
Abstract: In Uganda, using survey data, estimates of under-five mortality have only been feasible at national and regional levels. None exist at decentralized levels of governance like district level. Using small area techniques in Hierarchical Bayes framework, reliable estimates of relative risk of under-five mortality can be derived up to district level. The modeling approach have an added advantage over the commonly used Standardized Mortality Ratios (SMR) by estimating under-five disease risk for a particular district and smoothening using adjacent district estimates. The main objective of this study was to obtain estimates of under-five mortality that could be useful in targeting interventions at a district level and help to address the high burden of under-five mortality in Uganda.

16 June 2011

Presenter: Daniel Kadobera

09 June 2011

No meeting this week

02 June 2011

Presenter: Ernest Mwebaze
Causal Relevance Learning for Robust Classification under Interventions
The paper being presented is here:

31 March 2011

Presenter: Fred Kiwanuka

24 March 2011

Presenter: Guy Acellam
Mobile tools useful for AI projects: ODK Collect and OpenCV

17 March 2011

Presenter: George Okori

10 March 2011

Presenter: William Senfuma
Meta-Learning in Causal Structure Discovery

03 March 2011

Presenter: John Quinn

24 February 2011

Presenter: John Quinn

02 December 2010

Presenter: John Quinn

25 November 2010

Presenter: Fred Kiwanuka
Automatic Attribute Filter Parameter Selection For Blood Vessel Enhancement
The desirable property of connected attribute filters in identifying and extracting connected components without distorting their borders and no emergence of new ones has contributed to their popularity in image analysis. This paper explores novel, simple, fast and automated methods of extracting features for connected attribute filters based on image segmentation, thresholding and data clustering techniques. The performance of the techniques on blood vessel extraction is evaluated against the manual exhaustive user search method of computing the best threshold.

18 November 2010

Presenter: Richard Ssekibuule
Robust Collaborative Filtering

28 October 2010

Presenter: Kevin Leyton-Brown

21 October 2010

Presenter: Martin Mubangizi
Prediction of number of disease cases using dynamical Bayesian networks

14 October 2010

Presenter: Peter Nabende
Title: Applying edit distance based Dynamic Bayesian Network models in Transliteration identification
Transliteration identification involves the search for ‘new’ equivalent named entities between two different natural languages with the aim of building bilingual lexicons that are in turn expected to improve performance in major Natural Language Processing (NLP) applications such as Machine Translation (MT) and Cross Language Information Retrieval (CLIR). We explore the widely applied graphical models approach of Dynamic Bayesian Networks (DBNs) to transliteration identification in which we particularly evaluate the application of two edit distance based DBN methods. Both DBN methods model string similarity as a cost incurred through a series of edit operations. The first method implements Pair Hidden Markov Models (Pair HMMs) (Mackay and Kondrak, 2005) while the other implements DBN model classes that were initially introduced by Filali and Bilmes (2005) in their work on pronunciation classification. We modify both methods to specify DBN models that represent factorizations which we hypothesize could affect transliteration identification quality. We first test the DBN models on an experimental setup of English-Russian transliteration. We then test the Pair HMM approach on an English-Swahili Wikipedia dataset, and on the 2010 transliteration mining shared task datasets for the English-Tamil, English-Hindi, and English-Russian language pairs. High transliteration identification accuracy is obtained for the English-Russian experimental setup; however, a check on the output from DBN models associated with the two methods suggests that there can be improved transliteration identification quality through model combination. Although English and Swahili use almost a similar writing system, relatively low transliteration identification quality results from the Pair HMM approach on the English-Swahili Wikipedia data. For the other Wikipedia datasets, we get comparable transliteration identification quality to state of the art approaches despite using the Pair HMM in its ‘basic’ form.

07 October 2010

Presenter: Kevin Leyton-Brown (University of British Columbia)
SATzilla, SATenstein and Hydra: Attacking SAT with Automatic Design Patterns
Designing high-performance algorithms for computationally hard problems is a difficult and often time-consuming task. In this talk, we will consider three interrelated ``design patterns'' for expressing an algorithm design task as an optimization problem and solving it automatically. Because these optimization problems depend on the data upon which an algorithm is run, they can be understood as machine learning algorithms. While this talk focuses on the satisfiability problem (SAT), the methods it describes are general, and apply to any high-variance, computationally-intensive problem domain.
The first automatic design pattern stems from the observation that there is no single “dominant'' SAT solver; instead, different solvers perform best on different instances. Rather than following the traditional approach of choosing the best solver for a given class of instances, we show that it is possible to make this decision online on a per-instance basis. Specifically we describe SATzilla, an automated approach for constructing per-instance algorithm portfolios for SAT that use so-called empirical hardness models to choose among their constituent solvers. Our experimental results show that SATzilla outperforms its constituent algorithms. Indeed, twice in a row, it has also won medals in five of the nine categories in International SAT Competition.
Automatic design can also help when there does not exist a strong set of candidate algorithms among which to select. We first introduce a generalized, highly parameterized solver framework, dubbed SATenstein, that includes components gleaned from or inspired by existing high-performance SLS algorithms for SAT. The parameters of SATenstein control the selection of components used in any specific instantiation and the behaviour of these components. SATenstein can be configured to instantiate a broad range of existing high-performance SLS-based SAT solvers, and also billions of novel algorithms. We used an automated algorithm configuration procedure to find instantiations of SATenstein that perform well on several well-known, challenging distributions of SAT instances. Overall, we consistently obtained significant improvements over the previously best-performing SLS algorithms, despite expending minimal manual effort.
SATzilla and SATenstein each have both benefits and drawbacks. The latter has the advantage of requiring virtually no domain knowledge, but produces only a single solver; the former exploits per-instance variation, but must be given a set of relatively uncorrelated candidate solvers. In the last part of the talk I will introduce Hydra, a novel technique for combining these two methods, thereby realizing the benefits of both. Hydra automatically builds a set of solvers with complementary strengths by iteratively configuring new algorithms. It is primarily intended for use in problem domains for which an adequate set of candidate solvers does not already exist. Nevertheless, we tested Hydra on a widely studied domain, stochastic local search algorithms for SAT, in order to characterize its performance against a well-established and highly competitive baseline. We found that Hydra consistently achieved major improvements over the best existing individual algorithms, and always at least roughly matched—and indeed often exceeded—the performance of the best portfolios of these algorithms.

30 September 2010

Presenter: Ernest Mwebaze
Automated Vision Based Diagnosis of Cassava Mosaic Disease
(presentation from the Workshop on Data Mining in Agriculture 2010, Berlin)

23 September 2010

Presenter: Martijn Wieling (University of Groningen)
Hierarchical bipartite spectral graph partitioning to cluster dialect varieties and determine their most important linguistic features

16 September 2010

Speaker: Reading group
Discussing Barinova et al, "On Detection of Multiple Object Instances using Hough Transforms".

09 September 2010

Presenter: George Okori
Improved specificity famine prediction with satellite observation data

02 September 2010

Presenter: John Quinn
Spatiotemporal models for disease rate prediction with remote sensing data

08 July 2010

Group meetings in recess until August.

01 July 2010

Presenter: Jennifer Aduwo
Automated Vision-Based Diagnosis of Cassava Mosaic Disease
Cassava Mosaic Disease (CMD) has been an increasing concern to all countries in sub-Saharan Africa that depend on cassava for both commercial and local consumption. Information about the country-wide spread of this disease is difficult to obtain due to logistics and human resource issues in these countries. The objective of this study was to assess the feasibility of automated computer vision based diagnosis of CMD. Images of healthy and CMD-infected cassava leaves were taken at Namulonge Crop Resources Research Institute, Uganda. We performed classification on these images based on shape and colour features, using a set of standard classification methods (na\"ive Bayes, two-layer MLP networks, support vector machines, k-nearest neighbour and divergence-based learning vector quantization). We find near-perfect classification to be attainable for leaf images captured under ideal conditions, and outline a method for performing this classification on natural, cluttered images taken in situ.

24 June 2010

Presenter: Frida Katushemererwe
Intelligent systems for Bantu language learning with specific reference to Runyakitara

17 June 2010

Presenter: John Quinn
Updates on automatic malaria diagnosis with digital microscopy

10 June 2010

Group discussion: transfer learning

03 June 2010

Speaker: Jos Roerdink,
Scientific Visualization and Computer Graphics group,
Johann Bernoulli Institute of Mathematics and Computing Science
University of Groningen

Visualization is becoming ever more important in many areas, as it generically contributes to the interpretation of data which are of high dimension and/or large size. Techniques from graph visualization are increasingly applied to represent, retrieve, display, and explore complex systems, such as biological networks, brain connectivity networks, or large software systems, either as traditional two-dimensional images or in three dimensions, using interactive displays and virtual environments. Emphasis is put on interactive manipulation of visualized structures by providing users with tools to search, reorganize, control the level of detail (pan and zoom), interrogate, and derive new useful information.

I will discuss some general requirements for such systems and briefly mention a number of systems which have recently appeared. Almost invariably current systems are limited in terms of interactivity, adaptivity of views, possibilities for collaborative work, knowledge representation, allowable model dynamics, and literature coverage. I will briefly outline some recent more generic approaches based on graph visualization frameworks. Also, I will discuss techniques for using virtual environments for interactive exploration and navigation of multidimensional data, as well as collaborative visualizations using touch-table displays.

Finally, a new approach to study complex systems will be mentioned, called Visual Analytics, which pursues the integration of visualisation with other analytical methodologies, such as statistics, pattern recognition,data-mining, and cognition. This approach may have potential impact on many areas, such as science and engineering, health care, personalized medicine, safety and education.

27 May 2010

Presenter: Anthony Gidudu
Machine learning techniques for land cover mapping

20 May 2010

Presenter: George Okori
Data mining for patterns in famine data

13 May 2010

Presenter: Jon Gosier, Appfrica
Natural language processing and the SwiftRiver project.

06 May 2010

Presenter: Florence Tushabe
The Colour Attribute Filter

29 April 2010

Presenter: Peter Wakabi-Waiswa
Association Rule Mining Using Evolutionary Computing

22 April 2010

Presenter: Ernest Mwebaze
Divergence Based Learning Vector Quantization

We suggest the use of alternative distance measures for similarity based classification in Learning Vector Quantization. Divergences can be employed whenever the data consists of non-negative normalized features, which is the case for, e.g., spectral data or histograms. As examples, we derive gradient based training algorithms in the framework of Generalized Learning Vector Quantization based on the so-called Cauchy-Schwarz divergence and a non-symmetric Renyi divergence. As a first test we apply the methods to two different biomedical data sets and compare with the use of standard Euclidean distance.

15 April 2010

Presenter: Rose Nakibuule
Distinguishing vehicles and clutter in videos of traffic

08 April 2010

Reading group: highlights of AI-D

Paper summaries from Ernest:
1. Reality Mining Africa
Shawndra Hill, Anita Banser, Getachew Berhan, and Nathan Eagle
2. Quantifying behavioral datasets of Criminal Activity
Jameson Toole, Nathan Eagle, and Joshua Plotkin
3. Machine Learning methods for Verbal Autopsy in Developing Countries
Sean T. Green and Abraham D. Flaxman

Paper summaries from John:
1. People, Quakes, and Communications: Inferences from Call Dynamics about a Seismic Event and its Influences on a Population
Ashish Kapoor, Nathan Eagle, and Eric Horvitz
2. Case for Automated Detection of Diabetic Retinopathy
Nathan Silberman, Kristy Ahlrich, Rob Fergus, and Lakshminarayanan Subramanian
3. Intelligent Heartsound Diagnostics on a Cellphone using a Hands-free Kit
T. Chen, K. Kuan, L. Celi, and G. D. Clifford

01 April 2010

Presenter: Florence Tushabe
The Color Attribute Filter

This talk will introduce a color filter that is based upon connected operators. Color is defined in a unique combination of luminance and saturation and then ordered with respect to a reference vector. We tested the performance of the proposed filter on six popular types of images and applied it for automatic traffic signs recognition. Our experiments show that all the color schemes tested performed very well with others giving AUC/ROC results of as much as 85%.

25 March 2010

No meeting this week

18 March 2010

Presenter: Michael Biehl, University of Groningen, The Netherlands
Matrix Relevance Learning: An application to tumor classification

Similarity based supervised learning techniques, such as Kohonen's Learning Vector Quantization, can be extended and largely improved by the introduction of an adaptive distance measure. Here, we consider metrics which are defined by means of one (gobal) or several (local) relevance matrices. These take into account the importance of single features as well as correlations of different features.

We introduce and discuss the method in terms of an example application in the medical domain, i.e. the classification of adrenal tumor based on urinary steroid excretion. Here, relevance learning provides a criterion for the selection of the most discriminative steroids. We demonstrate that classification with high sensitivity and specificity is possible when using a reduced panel of steroids. The ultimate goal is the development of an efficient practical diagnosis tool.

11 March 2010

Presenter: Ernest Mwebaze
Causal structure learning for famine prediction

We consider the problem of understanding the causal relationships between socioeconomic factors in a developing-world household and their risk of experiencing famine. We analyse the extent to which it is possible to predict famine in a household based on these factors, looking at a data collected from 5404 households in Uganda. To do this we use a set of causal structure learning algorithms, employed as a committee that votes on the causal relationships between the variables. We contrast prediction accuracy of famine based on feature sets suggested by our prior knowledge and by the models we learn.

04 March 2010

Presenter: John Quinn
Traffic flow monitoring in crowded cities

We show how standard approaches to automatic traffic monitoring with CCTV cameras can be made more robust by using probabilistic inference, and in such a way that we bypass the need for vehicle segmentation. Instead of tracking individual vehicles we treat a lane of traffic as a fluid and estimate the rate of flow. Our modelling of uncertainty allows us to accurately monitor traffic flow even in the presence of substantial clutter.

25 February 2010

Presenter: Martin Mubangizi
Introduction to Kalman filtering and its application to predicting disease rates.

18 February 2010

Presenter: Jennifer Rose Aduwo
Classification of cassava mosaic disease with neural networks.

11 February 2010

Reading group: We will be studying and discussing the following paper
Makkapati, V. and Rao, R.M. Segmentation of Malaria Parasites in Peripheral Blood Smear Images IEEE ICASSP 2009 (pp.1361-1364)

04 February 2010

Presenter: Joyce Nakatumba
Introduction to process mining (using data mining techniques such as decision tree learning to find patterns in business processes).

28 January 2010

Presenter: John Quinn
Dynamical models for cellular tower data analysis