Abstract: 

Rainfall estimation from satellite data is of great importance, especially in the African rainfed agriculture setting where gauge stations are sparse. However, most of the precipitation estimation products rely on the relationship between cloud-top brightness temperature and actual rainfall, assuming that precipitation originates from convective clouds with cold tops. This leaves most of them underestimating rainfall in areas with warm cloud tops while overestimating rainfall in areas of cold cloud tops. This presentation we will be on fit-for-purpose validation of four satellite/reanalysis rainfall products for nine stations in Zambia.

Abstract: 

In classical survival analysis one focuses on a single event for each individual, describing the occurrence of the event by means of survival curves and hazard rates and analyzing the dependence on covariates by means of regression models. In other situations, more than one type of event is of interest. Such situations, with more than one possible type of event for each individual, may conveniently be described by multistate models. In credit risk analysis, multistate models recognize that "default" is only one of the possible credit rating grades to which the loan contract could migrate over the planning horizon. As such, these models estimate the probability that the borrower’s credit quality deteriorates, including a change to default status.

In this talk, we will discuss the application of multi-state models in predicting transition probabilities for a microfinance institution in Ghana. The models consider the transitions of borrowers among different credit status categories, such as full monthly payment, partial monthly payment, no monthly payment, and default on loan, and can incorporate covariates that may influence the transition probabilities. We compare the performance of various multi-state models, including the Markov chain model and the Cox regression model, in predicting the transition probabilities. 

Abstract: 

Climate change is a big challenge in all continents including Africa. Africa is particularly vulnerable to climate change impacts. Some of these impacts include floods, droughts, and storms which are the main sources of agricultural risk. This study examines the climate pattern in selected African countries (Kenya,

Cameroon, Somalia, Tunisia, Burundi, Rwanda, Madagascar, and Zambia) for a period of 1905 to 2021. To detect climate change patterns, precipitation and

temperature are analyzed using persistent homology, which is the main tool in topological data analysis (TDA) that focuses on qualitative information known as

topological features of data. Results of the analyses show that there are changes in topological features during climate change. The study also found that the first persistence landscape (a functional topological summary of persistent homology) consistently detects climate change. This paper offers the possibility to develop an effective early detection system for climate change based on our approach. 

Abstract: 

Deep Learning models have shown high performance in medical image analysis, specifically for retinal vessel segmentation. However, the current neural network models have not been able to represent the uncertainties that exist in images. The use of Bayesian neural networks has been shown across several studies to estimate the uncertainty of deep learning models. In this study, we briefly introduce uncertainty estimation and evaluate the predictive performance of some uncertainty estimation methods on retinal segmentation. We also show that using the estimates of the network in the loss function does not necessarily improve the segmentation performance. We end the talk by briefly discussing our current approach to get a reliable estimate of the uncertainty in this task. . 

Abstract: 

Global health is a domain which presents significant and pressing large scale challenges, any incremental improvement may have significant benefit at a global scale. In this work we focus on an important challenge in the space, calibration for disease models, and show that Reinforcement Learning (RL) is an effective algorithm for calibration problems at a scale which traditionally applied Bayesian approaches struggle. This work uses synthetic data, so has access to ground truth parameters and it can be seen that RL learns different, arguably better information for different parts of the learning process. These exciting results set the foundation for deeper consideration of RL in this space. 

Abstract: 

In numerous applications such as weather monitoring, it is desired to estimate a spatial field by observing it at a discrete set of locations. Often these observations are obtained using sensors and they can be contaminated by noise and anomalous values. In this talk, we describe the use of Gaussian process models for anomaly detection when observations are from a spatial field. Gaussian processes provide a natural way to deal with the spatial correlation of sensor observations. We present a variational Bayesian approach to inference in this setting and demonstrate the performance of the algorithm in detecting blocked rain gauges using data from a large weather station network deployed in Africa.

Abstract: 

One of the major challenges of machine learning and obstacle to more general intelligence is so-called continual learning. This refers to the regime of learning tasks in sequence. Humans are very good continual learners. We can learn X and then learn Y without forgetting  X, and in some cases also transfer learned aspects of X to Y. On the other hand artificial neural networks suffer from catastrophic forgetting, where information relevant to previously learned tasks is overwritten by the information required to effectively learn later tasks. In this talk I will discuss recent work trying to understand how the similarity of tasks in a continual learning setting affects the amount of forgetting observed in artificial networks. 

Abstract: 

Abstract: It was recently shown that the quantum annealing paradigm could be emulated in a variational Monte Carlo (VMC) framework using autoregressive neural networks [1]. There, only stoquastic driver Hamiltonians were considered. Here, we leverage the fact that the VMC algorithm is inherently sign-problem-free to simulate quantum annealing with non-stoquastic drivers. We show that the variational quantum annealing method is able to capture the dynamics of of non-stoquastic Hamiltonians and can provide an advantage for annealing paths that are hampered by exponentially closing gaps [2].


[1] M. Hibat-Allah， E. M. Inack, R. Wiersema, R. G. Melko， J. Carrasquilla, Nature Machine Intelligence volume 3, pages 952–961 (2021) 
[2] Nishimori and Takada，Frontiers in ICT 4，2 (2017)

Abstract: 

This talk will address the data format and archiving/compession issues for the Square Kilometre Array (SKA). One of the major contributors to the large data volumes is the maximum baseline of an array since it sets how well the data needs to be sampled to avoid smearing and decorrelation of the signal. Now, the sampling rate is dependent on the baseline length as shorter baselines can be sampled a lot more coarsely which leads to smaller data volumes. In fact, baseline-dependent averaging (BDA) is an established technique for compressing radio interferometric data, however, this technique results in irregularly sampled data which while supported by the database format requires the data to be restructured in ways that reduce performance when processing. But more importantly, radio-astronomy software tools implicitly assume uniformly sampled datasets. Recent research in M. Atemkeng, in prep. in collaboration with the Wits University and SARAO; shows an approach that uses local  low-rank matrix approximation to achieve similar compression rates while significantly minimizing smearing. 

Abstract: 

We investigate the role of stochasticity in linear recurrent neural networks (RNN). Time-allowing, I will discuss one or both of the following aspects:

A) The role of the noise in the computations in guaranteeing the robustness of learning (Based on a joint work with W. Bartolomeus, S. Nestler & H. Rauhut).

B) The role of the random choice of the connectivity matrix of the RNN in guaranteeing its separation capacity (based on an ongoing work). 

Abstract:

Digital and data-driven agriculture using Artificial Intelligence (AI), Blockchain, big data analytics, remote sensing and Internet of Things (IoTs) in farming industry necessitates the integration of various tools across the entire agricultural production system, as well as the use of enhanced systems-based techniques (i.e. a System of Systems based approach (SoS)) to ensure that it is scalable, adaptive, and sustainable.  Water scarcity, pests and diseases, climatic uncertainty, natural disasters, and labor shortages are all putting strain on crop and livestock agricultural production systems in arid and semi-arid African countries.  AI can accelerate the process of creating region specific crops and livestock by harnessing the tools of machine learning to mine existing and new geospatial and temporal, genotypic and phenotypic datasets to analyze relationships and predict outcomes. AI and remote sensing can speed up the process of developing region-specific crops by using machine learning methods to mine existing and new geographic and temporal, genotypic and phenotypic datasets to assess relationships and forecast outcomes. Real-time adaptive decision support systems that incorporate several data streams are required to optimize operations ranging from analyzing animal health to planning future pasture usage. Furthermore, developing AI for large-scale applications that can provide better short- and long-term management and planning for regional water and crop management is crucial. 

This paper provides a data-driven techniques for improving agriculture productivity of small-holder farmers in Africa. We provide our experience of developing digital agriculture platforms, climate-smart agriculture and farming as a business approach. We present various techniques for improving food security and productivity of agriculture using key technologies including: (a) Internet of Things (IoT) based agricultural systems, (b) blockchain for food supply chain monitoring and management, (c) AI/ML for increasing agricultural productivity, and (d)  remote sensing techniques for digital agriculture (e.g., imagery treatment and prediction in mapping). 

Abstract:

A recent report issued by the WWF states that there has been a catastrophic decline in wildlife population in recent years. A large number of species are threatened with extinction due to a number of factors such as over-exploitation of resources, deforestation and climate change. Certain species have been placed on the IUCN Red List for several years, but further conservation efforts are still urgently required to ensure the survival of the remaining individuals. While it is true that the number of individuals in threatened populations is decreasing, there have been considerable conservation efforts to put a halt to this. Ecologists, researchers, and rangers closely monitor the populations, in certain cases by placing microphones and camera-traps into the environment and searching through the recorded audio/images for the species of interest. This non-invasive approach comes with a cost in that huge datasets of audio or images are produced and is difficult to manually process. This talk will discuss efforts in using machine learning to monitor certain species through the use of machine learning. The theme of this colloquium is: "Helping Africa Listen to Herself" where I will showcase some of our latest efforts as well as efforts by my students in producing machine learning models for conservation ecology. 

Abstract:

Foundation Models are a family of models that use unsupervised objectives to learn representations of data units. These models demand re training on mammoth volumes of data at scale and are adaptable to a wide range of downstream tasks. Research and Industry has already experienced tremendous and burgeoning development in the use of these models for natural language and image processing e.g., with the models such as  BERT, LaMDA, DALL-E , and GPT already showing viable products in the public domain. Nonetheless, the range of application possibilities for these models is enormous and can only be viewed as unexplored. Research begins to extend these models with concepts from physics and thermodynamics for application in healthcare, and discovery science - drug discovery, molecule discovery etc. We take a look into what foundation models are and their background but provide a brief peek into the applications.    

Abstract:

Visualising high-dimensional datasets is a crucial step in exploratory data analysis. However, there has been some controversy about the relation of the two most prominent non-linear visualisation methods, t-SNE and UMAP. They have seemingly unrelated objective functions and distinct key ingredients, producing qualitatively different embeddings.

We find the exact relation between UMAP and t-SNE and show that UMAP is essentially negative sampling applied to the t-SNE loss function. Our main insights are the derivation of UMAP’s true loss function and its connection to noise-contrastive estimation, which is used by the method NCVis to approximate t-SNE.

Our analysis shows that UMAP applies stronger attraction on neighbouring points than t-SNE and thus explains the more compact embeddings produced by UMAP. Varying the attraction strength further leads to a whole spectrum of neighbour embedding methods on which t-SNE, UMAP and NCVis are only three points.

Gauging attraction and repulsion has important practical consequences. On the one hand, stronger attraction can highlight continuous structures, like trajectories. On the other hand, more repulsion resolves discrete clusters better and thus shows the finer local structure of a dataset. We recommend looking at different points on the attraction-repulsion spectrum to observe how (possibly artifactual) structure arises and decays in order to faithfully explore high-dimensional datasets 

Abstract:

Most cancer patients die due to metastasis, and the early onset of this multistep process is usually missed by current staging tumour modalities [1]. Advanced technics exist to enrich disseminated tumour cells from patient blood and bone mallow as cancer progression marker. However, these cells present high heterogeneity [2], only some of them can exhibit stem cell phenotype and tumour development potential, other can have plasticity potential to reprogram into cancer stem cells [3]. So, detection and characterization is challenging due to lack of clear phenotypic markers [4]. Therefore, there is a critical need to find new ways to anticipate and predict metastasis development at an early stage of patient care. Cancer progression involves many cellular morphological effects, which have been revealed by biophysical studies [5]. The relevance of the characterization of cancer cells by their electromechanical phenotype is attested by reports pointing out their physical alteration as reduced deformability of circulating lymphocytes in the case of chronic lymphocytic leukaemia, and increased size and stiffness of labelled circulating tumor cells (CTC). The physical properties allow identifying different malignant breast epithelial cells lines by their viscoelastic behavior [6] or their electrical impedance [7]. Even though, the analysis capability of cancer cell by their physical characteristics has been demonstrated [5], the current state of the art is far from being relevant for clinical practices. This project aims to develop and push the concept of cell physical phenotyping to categorize disseminating cells, evaluate their metastatic potential towards cancer diagnosis and prognosis. We aim to combine MEMS (Micro-Electro-Mechanical Systems), biological assays, statistical learning ([8]) with following objectives:

• Physical phenotyping: A high-throughput method for physical characterization of cells

• Biological phenotyping: Characterization of cells to analyze their metastatic potential

• Modeling: Statistical functional methods for physical/biological cell classification

• Predicting the metastatic potential of cells

This work offers a complete mapping and correlation of different characteristics (biological, physical and genetic) of selected cells lines to model their metastatic potential for improved diagnostic and prognostic capabilities. It aims providing the required statistical/ML tools to predict metastatic potential of a cell by

means of physical properties; a method that is faster, cheaper and better suited for point-of-care applications. For data processing, the needed statistical and computational classification and prediction methods require training with large-scale, heterogeneous, continuous (functional) and spatial datasets including

high numbers of cells’ physical and biological properties. A very limited number of models are available for description, visualization, classification and prediction of quantitative data involving continuous (functional) heterogeneous data. Challenges are, in one hand, to use statistical (including functional

classification, regression methods) tools able to compute in a non-costly way correlation among huge amounts of data, for training and accuracy prediction. A main issue when analyzing our massive data is to use statistical tools including functional classification, regression methods) able to compute in a non-costly way correlation among huge amounts of data and the ultimate goal of predicting the metastatic potential of cells by physical characterization of cells.

The work is interdisciplinary (biophysicists, biologists with analysts and statisticians) and is in collaboration with University of Tokyo (Bio-MEMS technology in SMMiL-E ; Seeding Microsystems in Medecine in Lille – European-Japanese Technologies against Cancer) project: http://www.ircl.org/programme-smmil-e/).

References:

[1] Fidler IJ. Nat Rev Cancer, 2003, 3, 453–458.

[2] Punnoose EA, Atwal SK, Spoerke JM, Savage H, Pandita A, et al.,. PLoS One. 2010, 5:e12517.

[3] Lagadec C; Vlashi E; Della Donna L; Dekmezian C; Pajonk F. Stem Cell, 2012, 30(5), 833-44.

[4] Millner L, Linder M, Valdes R. Ann Clin Lab Sci. 2013, 43(3), 295–304.

[5] S. E. Cross, Y.-S. Jin, J. Rao, and J. K. Gimzewski, Nature Nanotech, 2007, 2, 12, 780–783.

[6] J. Guck, S. Schinkinger, B. Lincoln, F. Wottawah, S. Ebert, et al.,, Biophys. J, 2005, 88, 5, 3689–3698.

[7] A. Han, L. Yang, and A. B. Frazier, Clinical Cancer Research, 2007, 13, 1, 139–143.

[8] A. Bahram et al. MEMS, 2022, 2022, 317-320 

Abstract:

Progress toward the United Nations Sustainable Development Goals (SDGs) has been hindered by a lack of data on key environmental and socioeconomic indicators, which historically have come from ground surveys with sparse temporal and spatial coverage. Recent advances in machine learning have made it possible to utilize abundant, frequently-updated, and globally available data, such as from satellites or social media, to provide insights into progress toward SDGs. Despite promising early results, approaches to using such data for SDG measurement thus far have largely evaluated on different datasets or used inconsistent evaluation metrics, making it hard to understand whether performance is improving and where additional research would be most fruitful. Our contributions our  two-fold. First, we have compiled one of the largest ML-ready datasets for tracking progress towards UN SDGs, which we call SustainBench. This is a collection of  15 benchmark tasks across 7 SDGs, including tasks related to economic development, agriculture, health, education, water and sanitation, climate action, and life on land. It significantly lowers the barriers to entry for the ML community to contribute to measuring and achieving the SDGs and provides standard benchmarks for evaluating ML models on tasks across a variety of SDGs. Second, we have developed novel machine learning methods that have directly facilitated progress towards the SDGs, such as enabling targeted cash-transfer programs at the start of the COVID-19 pandemic. Nonetheless, there remains significant room for developing new AI methods for further accelerating progress towards the SDGs.

Abstract: 

This talk will address the data format and archiving/compession issues for the Square Kilometre Array (SKA). One of the major contributors to the large data volumes is the maximum baseline of an array since it sets how well the data needs to be sampled to avoid smearing and decorrelation of the signal. Now, the sampling rate is dependent on the baseline length as shorter baselines can be sampled a lot more coarsely which leads to smaller data volumes. In fact, baseline-dependent averaging (BDA) is an established technique for compressing radio interferometric data, however, this technique results in irregularly sampled data which while supported by the database format requires the data to be restructured in ways that reduce performance when processing. But more importantly, radio-astronomy software tools implicitly assume uniformly sampled datasets. Recent research in M. Atemkeng, in prep. in collaboration with the Wits University and SARAO; shows an approach that uses local  low-rank matrix approximation to achieve similar compression rates while significantly minimizing smearing.