The insufficiency of ground truth data is usually a barrier when researchers are trying to train machine learning models for real-world applications. In this talk I will be discussing some of the computer vision applications I have worked on in the past, discussing some interesting techniques as well as lessons learned while creating ground truth labels in order to answer research questions around ways to study the effects of Spatial Apartheid in South Africa.

On-device Machine Learning applications provide a wealth of opportunities for sensing and analytics, particularly when cloud connectivity is not always readily available. Making these applications more energy-efficient and private can reduce their reliance on batteries and/or excessive data collection. Solutions in this space would have significant implications for a new generation of sensing and monitoring applications for environmental monitoring, population-wide analytics, scientific exploration, and climate/weather prediction. In this talk I will provide an overview of recent attempts in this space, and challenges ahead for providing reliable, secure, and private client-side applications.

Crop type mapping from satellite images is an essential input for food security monitoring systems. Many approaches focus on mapping crop types based on a full time series of a growing season. However, a variety of use cases require predictions already during the growing season which can be technically challenging. In this paper, we experiment with Sentinel-2 and Planet Fusion data to explore their potential for early season crop type classification at different points in the season. We use high-quality field collections from Germany and South Africa as reference data and find that daily revisit times can be advantageous but are no silver bullet for early season classification of crops.

In this work, we share our methodology and findings from applying named entity recognition (NER) using machine learning, to identify behavioural patterns in transcribed family planning client call centre data in Nigeria based on the Fogg’s model. The Fogg Behaviour Model (FBM) describes the interaction of three key elements Motivation (M), Ability (A), and a Prompt (P) and their interaction to produce behavioural change. This work is part of a larger project that is focused on practical application of artificial intelligence to analyse and derive insight from large scale data call centre data. The entity recognition model called Fogg Model Entity Recognition(FMER) was trained using spaCy, an open source software library for advanced natural language processing on a total of 11510 words and F1 score of 98.5%.

In India, the majority of farmers are classified as small or marginal, and their livelihoods are vulnerable to climate risk. One approach to mitigating this problem is via drip-irrigated greenhouses. We aim to build an optimization-based decision support tool that provides crop planning advice to farmers in conjunction with a nonprofit in India. We propose a Markov decision process approach for this low-resource sector, and discuss key evaluation metrics for the design and implementation phases of our project.

This paper describes the initial activities of the X* Project, a COVID-19 motivated research effort to design a system for the early prediction of respiratory insufficiency via audio analysis, describing the research motivation, its organization in research lines, the initial results obtained in those lines and a preview of the future steps in this research project.* Project name omitted due to blind revision

In text classification tasks, fine tuning pretrained language models like BERT andGPT-3 yields competitive accuracy; however, both methods require pretraining onlarge text datasets. In contrast, general topic modeling methods possess the ad-vantage of analyzing documents to extract meaningful patterns of words withoutthe need of pretraining. To leverage topic modeling’s unsupervised insights ex-traction on text classification tasks, we develop the Knowledge Distillation Semi-supervised Topic Modeling (KDSTM). KDSTM requires no pretrained embed-dings, few labeled documents and is efficient to train, making it ideal under re-source constrained settings. Across a variety of datasets, our method outperformsexisting supervised topic modeling methods in classification accuracy, robustnessand efficiency and achieves similar performance compare to state of the art weaklysupervised text classification methods.

Over the recent years, there has been excitement about the extraordinary opportunities that artificial intelligence may offer in tomorrow’s healthcare. Given the potential of AI technology in facilitating the quantification of large and complex datasets, medical imaging has witnessed rapid and revolutionary developments. However, a limitation of current AI developments for medical imaging is that they have overwhelmingly, and almost entirely, targeted imaging applications in highincome settings. Hence, it is important to promote and accelerate the development of trustworthy and accessible AI solutions for medical imaging in low-to-middle income countries –an urging need to advance global healthcare. This paper describes the authors’ experience and initiatives in promoting AI for medical imaging on the African continent, by Africans for Africans. First, the paper will discuss MICCAI 2024, the first international conference on medical image computing and computer assisted intervention that will be taking place on the continent. Subsequently, we will present AFRICAI, a new African network dedicated to research, education and cooperation in the field of AI in imaging and radiology. With this paper, we hope to raise awareness about these initiatives and attract more collaborators, promote new research and innovation in the field to address Africa-specific healthcare challenges, and encourage similar initiatives for promoting practical AI solutions for developing countries in Africa and beyond.

Identifying change in remote sensing observations - such as satellite images - is an important step for many applications. In practice, the aim is usually not to find all differences, but rather only specific types of change, such as urban development, deforestation, or even more specialized categories like road construction. However, often there are no large public data sets available for very fine-grained tasks or non-standard usecases that might, for example, occur in developing countries, and to collect the amount of training data needed for most supervised learning methods is very costly and often prohibitive.For this reason, we formulate the problem of few-shot filtering, where we are provided with a relatively large change detection data set and a few instances of one particular type of change that we try to “filter out” of the learned changes. This would enable stakeholders to work with the available resources and adapt them to their individual needs, for a wide range of applications.

Poverty maps—spatial representations of economic wealth—are essential tools for governments and NGOs to adequately allocate infrastructure and services in places in need. They also help to better understand social phenomena such as human mobility and segregation, and environmental problems induced by urbanization. Traditionally, such maps are inferred from Census and survey data, which are expensive and collected occasionally; thus, they commonly provide outdated and low-resolution socioeconomic information, especially in developing countries. Remotely sensed data combined with advanced machine learning methods provided a recent breakthrough in poverty map inference. However, these models are not optimized to produce accountable results that guarantee accurate predictions to all sub-populations including the rich and the poor or the urban and rural divide. In this paper, we touch upon the opportunities that multimodal data can offer to solve these issues, as well as the challenges of working with noisy, biased and sparse datasets for predicting high-resolution poverty maps in Sierra Leone.

Breakthroughs in the use of AI for Diabetic Retinopathy(DR) diagnosis, have made headway in making DR treatment more accessible but image and camera variability significantly affects the reproducibility of these machine learning algorithms. In an effort to improve the reproducibility of ML algorithms, we attempt to build a retinal image processing pipeline to quantify image quality taking into account luminance and blurriness, discarding poor quality images based on these metrics. Our pipeline further standardizes all images by cropping and resizing. To test the impact of our processing pipeline, we document the results of a 5-fold cross validation with and without the pipeline. Running images through the pipeline shows an increase in AUC performance attributable to an increase in image quality.

Till now, attention-based models have been used with great success in the keyword spotting problem domain. However, in light of recent advances in deep learning, the question arises whether self-attention is truly irreplaceable for recognizing speech keywords. We thus explore the usage of gated MLPs---previously shown to be alternatives to transformers in vision tasks---for the keyword spotting task. We provide a family of highly efficient MLP-based models for keyword spotting, with less than 0.5 million parameters. We show that our approach achieves competitive performance on Google Speech Commands V2-12 and V2-35 benchmarks with much fewer parameters than self-attention-based methods.

In this talk, we highlight the challenges of three practical machine learning(ML) projects conducted by researchers in Argentina. First, we present our collaboration with the Latinamerican and the Carribean hub of the Feminist AI Research Network (aka, fr). Second, we talk about a project that uses Natural Langugae Processing for the de-identification of electronic health records in collaboration with La Rioja province in Argentina. Finally, we will describe a project applied to agroecological vegetable gardens in Patagonia. This talk will sythesize reflections about the practicalities and the ethical considerations of applying ML. We will also include our lessons learned from collaborating with other researchers based in Latin America and the Carribean.

The human upper nasal airway is an anatomical structure with a complex geometry that performs essential functions required by the rest of the respiratory system. An accurate and precise segmentation process that captures its intricate shape and variability becomes indispensable to fully understand its performance under different circumstances and to study its anatomy from multiple perspectives. As currently performed, the manual or semi-automatic segmentation process for these structures is extremely time-consuming, may demand extensive manual post-processing steps to correct over-or under-segmentation, and is subject to considerable intra- and inter-operator variance. Further, in developing countries, healthcare institutions modernize their medical imaging devices at different rates; thus, specialists and proposed solutions have to deal with a wide range of image characteristics and quality variability to execute their diagnostics. In this paper we develop an automatic segmentation strategy for the human upper nasal airway,based on a deep convolutional network trained with >3000 CT scans acquired from different devices of a national hospital in Argentina, Hospital Italiano de Buenos Aires (2010). This process achieves a remarkable preliminary results with a low error rate (0.07%) and an acceptable similarity score (86.9%).

The cropping up of illegal landfills across the world and more so in developing countries has become a prevalent malice. It has become essential to come up with innovative solutions to this challenge, one that goes beyond detecting these spots manually. In recent times, remote sensing along with deep learning and computer vision has been applied to solve many problems. Through this paper, we would like to present a very high resolution Landfill dataset created from satellite images and demonstrate that by applying suitable deep learning methods, landfills can be detected even with a constrained and limited dataset. Through this paper we also make the Landfill dataset available to the research community

As acute climate change impacts weather and climate variability, there is increased demand for robust climate impact model predictions from which forecasts of the impacts can be derived. The quality of those predictions are limited by the climate drivers for the impact models which are nonlinear and highly variable in nature. One way to estimate the uncertainty of the model drivers is to assess the distribution of ensembles of climate forecasts. To capture the uncertainty in the impact model outputs associated with the distribution of the input climate forecasts, each individual forecast ensemble member has to be propagated through the physical model which can imply high computational costs. It is therefore desirable to train a surrogate model which allows predictions of the uncertainties of the output distribution in ensembles of climate drivers, thus reducing resource demands.This study considers a climate driven disease model, the Liverpool Malaria Model (LMM), which predicts the malaria transmission coefficient R0. Seasonal ensembles forecasts of temperature and precipitation with a 6-month horizon are propagated through the model, predicting the distribution of transmission time series. The input and output data is used to train surrogate models in the form of a Random Forest Quantile Regression (RFQR) model and a Bayesian Long Short-Term Memory (BLSTM) neural network. Comparing the predictive performance, the RFQR better predicts the time series of the individual ensemble member, while the BLSTM offers a direct way to construct a combined distribution for all ensemble members. An important element of the proposed methodology is that accounting for non-normal distributions of climate forecast ensembles can be captured naturally by a Bayesian formulation.

Hate speech and misinformation, spread over social networking services (SNS) such as Facebook and Twitter, have inflamed ethnic and political violence in countries across the globe. We argue that there is limited research on this problem within the context of the Global South and present an approach for tackling them. Prior works have shown how machine learning models built with user-level interaction features can effectively identify users who spread inflammatory content. While this technique is beneficial in low-resource language settings where linguistic resources such as ground truth data and processing capabilities are lacking, it is still unclear how these interaction features contribute to model performance. In this work, we investigate and show significant differences in interaction features between users who spread inflammatory content and others who do not, applying explainability tools to understand our trained model. We find that features with higher interaction significance (such as account age and activity count) show higher explanatory power than features with lower interaction significance (such as name length and if the user has a location on their bio). Our work extends research directions that aim to understand the nature of inflammatory content in low-resource, high-risk contexts as the growth of social media use in the Global South outstrips moderation efforts.

Over 10% of the global population is affected by chronic kidney disease (CKD) and those without preventative care and early intervention are the worst impacted. Many childhood precursors to CKD such as hydronephrosis (HN) continue to be detected and treated late in low- and middle-income countries where prenatal and early-life ultrasound is less common. Artificial intelligence-based technology holds promise for improving some of this detection and treatment. In this work, we explore the promise and pitfalls of transferring an AI-based tool for early HN detection in pediatric ultrasound from Canada, where it was developed and validated, to South Africa. We explore challenges and opportunities at the health-system-, institutional-, and provider-levels that using this technology in this new clinical environment involve. Our investigation is performed through interviews with clinicians at various levels, locations, and in different specialties. We find that the context of our tool's use will change in terms of both clinicians and patients, as the users of our tool in South Africa will have less access to pediatric sonography expertise and, for related reasons, patients will tend to be older when they receive an ultrasound imaging. These differences imply the need for adaptation of both the underlying algorithm as well as the tool's interface for successful deployment in this setting. However, given these revisions, the clinicians interviewed are eager for AI-based assistance in caring for these patients earlier and more effectively and believe algorithms of this kind will be useful for improving care.

Nursing is centered on the theory and practice of caring. Caring goes well beyond caring to bring patients to health. It starts with respect and equanimity that is achieved by respecting the values of all patients, even if they are different from yours, and treating patients as they are. Taking inspiration from nursing, in this talk, I will discuss how a carative approach to AI should start with the real-world problem as experienced by the most vulnerable, listening to and understanding their values, meeting them where they are, working toward a solution to their problem all the way to the end (even if it involves a lot of grunt work and doesn't lead to an ICLR paper), and conducting a qualitative assessment of the entire solution by interviewing the affected communities.