AMLD Africa 2024

Welcome and Opening Remarks

Community Presentations

Presentation of the communities:
13:45 - Mawazoo Institut
13:51 - Masakhane
13:57 - Africlimate
14:03 - Kenya AI
14:09 - Women in Machine Learning & Data Science
14:15 - Tanzania AI
14:21 - Ken corpus
14:27 - Zindi
14:33 - Indaba
14:39 - AI community of practice

Speed Dating

Responsible AI in Africa, challenges and opportunities

Edtech AI in Africa from founders

With Kiko Muuo, Mumbe Mwangangi, Ngugi Karega and Felix Malombe.
This panel discussion brings together three founders of EdTech startups in Africa to delve into the intersection of artificial intelligence (AI) and education. The discussion will explore how AI is transforming the landscape of education in Africa, from personalized learning experiences to addressing accessibility and equity challenges. The founders will share insights into their innovative approaches, challenges faced, and the impact of their solutions on learners and educators across the continent. Topics to be covered include the role of AI in enhancing student outcomes, the importance of data privacy and ethical considerations, and the potential of AI to bridge educational gaps and empower marginalized communities. Join us as we uncover the transformative power of AI in shaping the future of education in Africa.

Interlude: Rhythms of the continent

Harnessing the Power of Artificial Intelligence to end new HIV Infections among the Kenyan Youth

Closing

Africlimate workshop - Towards Building an African AI and Climate Science Research Community

[Intermediate level]
With Ibrahim Yusuf, John Bagiliko, Grace Wangui, Rendani Mbuvha and Santiago Hincpie Potes.
The workshop is designed to tackle the urgent issue of climate change in Africa by using the power of Artificial Intelligence (AI) in climate science. Its goal is to gather a diverse group of experts, researchers, and practitioners to discuss AI solutions specifically designed for the climate challenges in Africa. It will highlight the work of the AfriClimate AI community in climate AI research across the continent.
The agenda includes:
• A keynote speech by Prof Tshilidzi Marwala, Rector of the United Nations University and UN Undersecretary General.
• A panel discussion.
• Several spotlight talks by members of the AfriClimate AI community (3-4 talks).
• Two tutorials: one on creating a climate data pipeline and another on deep learning for climate modeling using GraphCast.
The workshop offers an introduction to climate modeling and its challenges, hands-on building of climate data pipelines, and a discussion on current opportunities with domain experts.

Getting started with Polars (when Dataframes get a little too fast)

[Beginner level]
If you’ve ever wondered how to speed up data analysis in your Python DataFrames, then considerer Polars, a popular DataFrame library written from the ground up in Rust. https://github.com/pola-rs/polars This workshop will start with the basics of Polars, along with a comparison to Pandas DataFrame. Participants will walk through Python code exploring functions and features of Polars, for example load and transform data from CSV, Excel, Parquet, IPC or a database connection, perform data analysis in parallel and prepare data for machine learning pipelines. This is a hands-on workshop so please bring your laptops and you’re excepted to have some basic background with programming/Python and Jupyter notebooks. If you’ve been curious about DataFrames in Python and how to get started then do feel free to join us too! Participants will learn how to get started with Polars, the syntax and how to break down problems into code.
Prerequisites:
1. Must be familiar with Python
2. Have used Pandas before even if it is in basic use cases
3. Must bring laptop with Python 3 installed

Registration

Opening Ceremony

Join us for the AMLD Africa opening ceremony which sets the stage for an enriching journey of exploration and collaboration. Speakers: Omolola Omole-Odubekun Prof. Salathé, Founder and Co-organizer of AMLD EPFL Mohamed Ali Dhraief and Kamil Seghrouchni, Executive Committee of AMLD Africa

AGI4Health: Building the world’s first true Generalist AI Doctor

Towards Autonomous AI Agent for Drug Discovery

The talk will delve into how artificial intelligence is reshaping drug discovery. It will discuss the current limitations and challenges in fully leveraging AI to industrialize this field and enable widespread access to treatments. It will spotlight tools like LOWE (LLM-Orchestrated Workflow Engine) for their role in optimizing complex workflows and expediting drug development. Additionally, the talk will emphasize the importance of selecting appropriate inductive biases in designing ML architecture and incorporating human insights to notably improve predictive accuracy to drive innovation on an unprecedented scale.

Coffee Break

Context-adapted Large Language Models to Improve Health Access in Africa

Large language models (LLMs) have the potential to democratize access to medical knowledge. Unfortunately, the enormous potential of these models is either locked behind commercial/research licenses, in violation of privacy regulations, limited in scale, or not generalizable to underserved populations and resource-limited settings. To address this issue, we developed Meditron-70B, currently the world’s best-performing fully open-source chatbot for medicine, trained on carefully curated clinical practice guidelines from diverse settings. However, the performance of these chatbots is commonly measured on medical exam questions, which does not adequately evaluate real-world clinical utility and safety. In this talk, I introduce Meditron and show how we are crowdsourcing incentivized expert evaluations that is putting Meditron to the test. I introduce the MOOVE (Massive Online Open Validation and Evaluation) platform that allows doctors to validate the real-world performance of Meditron in terms of helpfulness, harmlessness, bias, trust, and safety. In return for this rigorous validation, participants can get their own chatbot, adapted to their preferences and specialty.

Evaluating Large Language Models for Safe and Effective Use in Healthcare (Large language models)

Evaluating Large Language Models for Safe and Effective Use in Healthcare Large language models (LLMs) like ChatGPT show promise for transforming healthcare delivery through natural language processing applications. However, concerns exist regarding the potential for generating misinformation, lacking transparency, and perpetuating biases. This presents challenges for the safe and ethical adoption of LLMs in clinical settings. A comprehensive framework is proposed to evaluate LLMs’ utility and governance when applied in healthcare. The framework incorporates natural language processing metrics along with assessment of translational value across capability, utility, and adoption dimensions. Governance components emphasizing fairness, transparency, trustworthiness, and accountability provide oversight for responsible LLM use. Together, these layers allow thorough, multifaceted analysis of benefits and risks to guide appropriate LLM adoption while ensuring patient safety. The suggested framework aids stakeholders in realizing LLMs’ potential while proactively addressing limitations, supporting informed decisions around their healthcare integration.

Harnessing hidden signals in medical images

Recent studies have shown that both biologic and non-biologic disease-based characteristics can be predicted from medical imaging. For instance, self-reported race, sex, and age can be predicted from chest X-rays, retinal scans and other imaging modalities, as well as disease conditions, such as ICD code diagnoses. Additional research has indicated that disease risk, such as the risk of breast cancer, can be predicted from medical imaging with better performance than clinical and traditional scoring systems like the Tyrer-Cuzick and Gail breast cancer risk prediction. These models prove to be powerful even when they lack high precision labels from radiologists. However, these image models face challenges due to the inadequacy of existing explanatory techniques. Furthermore, given the known issue of shortcut learning causing bias, there is increased concern over the use of image-only models. Conversely, if properly leveraged, image-only models can be successful, particularly for opportunistic screenings and mining information for population health, even if their initial intent was not for subsequent use. I will discuss image-only models and their methodologies that have demonstrated superior performance over non-traditional imaging-only models, as well as the challenges and limitations of scaling imaging for precision medicine, specifically shortcut learning and limited explainability techniques.

Deploying Models in Production: Optimization, Speed, and Scalability

While models can generate remarkable predictions and valuable outputs, their true worth lies in their accessibility to end-users. Manual deployment processes are often cumbersome and costly, hindering the efficient delivery of models for end-users. Automated deployments offer a promising alternative that ensures rapid delivery of updated models and numerous other benefits. Efficient deployment mechanisms must prioritize resource utilization, speed, and scalability to ensure robust and enriching user experiences. Techniques like pipelines, containerization, microservice architecture and auto-scaling infrastructure are some of the explorable options in optimizing deployments. In this workshop, we explore an industry-centric perspective on model deployment and system architectures, drawing from minoHealth AI's deployment practices, to showcase the efficiency gains achievable by doing it right. We also provide insights into optimizing inference and generation processes and scalability considerations for deployments.

Interlude: Rhythms of the continent

Novartis Presentation - Health Tech Innovation Hub

Novartis is a pharmaceutical company reimagining medicine to improve and extend people’s lives. Within Novartis, we have Global Health & Sustainability unit, which strives to ensure everyone can benefit from Novartis innovation regardless of where they live or their socio-economic situation; and to maximize our social impact as a sustainable business. Sub-Sahara Africa is also part of Global Health & Sustainability unit within Novartis. As we aim to continue maximizing our impact in Sub-Sahara Africa, one of the innovative ideas is the HealthTech Hub Africa to Accelerate health tech innovation (https://thehealthtech.org/). HealthTech Hub Africa addresses the same challenge we are tackling with like-minded partners such as AMLD Africa in health care.

Applying AI to enable Biological discoveries

AI holds much promise for the future of healthcare and biomedical sciences. Perhaps the most striking and most important feature of modern AI technology is how much it differs from biological brains. A modern AI architecture is as distant from the human brain, that plane is from a bird. In this definitive distinction lies the potential for great cognitive and computational complementarity between AI and Researchers. In this presentation I will highlight some examples of how applied AI was used to enable fundamental discoveries in biology.

The AI Doctor Will See You Now: Navigating AI's Ethical Impact on Healthcare Equity & Inclusion

Imagine a future where AI supports human doctors, enhancing diagnoses, making treatments more effective, and saving countless lives. Exciting, right? But there's a big question mark: Will these AI advancements benefit everyone equally, or will they leave some behind? In this talk, we'll dive into the exciting yet challenging world of AI in healthcare. We'll explore why it's crucial to ensure AI is fair and inclusive, touching on the pitfalls like bias and access disparities that could sideline underrepresented communities. But don't worry, it's not all doom and gloom! This talk will also focus on how to overcome those challenges to make sure AI in healthcare benefits all - no matter where they are in the world. Join this session to discover how we can navigate the ethical maze of AI together and build toward a future where using AI truly enhances healthcare for everyone.

Panel Discussion

With Emmanuel Noutahi; Mary-Anne Hartley, Kamil Seghrouchni; Judy Gichoya; Therence Bois; Join us for a captivating panel discussion featuring African AI founders sharing their experiences, insights, and advice. Gain valuable perspectives on market trends, key considerations, challenges, and future outlooks in the AI landscape. This concise session promises to be both informative and inspiring, offering practical wisdom for aspiring entrepreneurs in Africa’s burgeoning AI ecosystem.

Day 1 - Closing Remarks

An artificial intelligence approach to malaria case detection in Siaya, Kenya

The study investigates the potential of wearable technology for early malaria detection in Siaya, Kenya. Utilizing wearables to monitor vital signs, the research employs machine learning models such as GAN with LSTM, NCP, and WGAN for data analysis. Initial results show promise for early detection in low-resource settings. Challenges include data variability and the need for larger datasets. Future work will focus on model refinement and broader population studies to develop a scalable tool for resource-limited health systems.

A Study of Acquisition Functions for Medical Imaging Deep Active Learning

In the real world, especially in the medical imaging context, data scarcity and limited labeled data are recurrent and frequent problems. This is very often a bottleneck to high-performance of recent Deep Learning approaches that are very data-hungry. In this work, we show that active learning could be very effective in data scarcity situations, where obtaining labeled data is expensive. We compare several acquisition functions (AF) such as BALD, MeanSTD, and MaxEntropy on the ISIC 2016 Melanoma detection dataset, explore the impact of selecting either the most or least uncertain samples, and leverage the effect of acquired pool sizes on the performance of the model. Our results on the Melanoma detection test set, demonstrate that uncertainty is useful to the Melanoma detection task and that it is more beneficial to select the most uncertain pool samples. These results suggest that active learning could be very useful for medical imaging tasks (in particular) and more generally in low-resource settings.

Algorithms to improve fetal-maternal health using magnetic resonance imaging

The placenta is an organ that connects the fetus to the maternal blood system to provide oxygen and nutrients to enable fetal growth. Placental dysfunction can affect both the child’s health and the mother’s health, so there is a critical need to assess placental health and function in vivo, or during pregnancy. Magnetic resonance imaging (MRI) in pregnancy has emerged as a promising imaging technique to quantify placental function and study health conditions in pregnancy. However, interpretation and analysis of the MRI scans is challenging as the images are low resolution, have low SNR and contain significant artifacts caused by fetal motion. Consequently, algorithmic techniques are necessary to extract relevant and meaningful signal in the placenta to quantify health and function. In this talk, I will provide an overview of several problems in placental image analysis including segmentation, registration, and visualization. I will describe some of our group’s work on these topics and highlight open challenges. I will conclude by discussing the potential of this work in clinical research studies to improve fetal-maternal health.

Infant Cognitive Monitoring Using Computer Vision

With the help of modern technology in telemedicine, significant progress has been achieved in the remote diagnosis and treatment of patients. These solutions rely on telecommunication technology to transmit data; however, their applications have been confined to measurements that can be taken mechanically such as monitoring blood pressure, breathing rate and heart rate. Imaging solutions are rarely used for Remote Patient Monitoring (RPM) because they require skills and present high costs. In marginalized areas that are in dire need of RPM, these challenges present serious barriers. This research presents a prototype that automates RPM through computer vision. The solution makes use of object tracking and object detection techniques to process images and videos taken by ordinary digital cameras found in mobile phones to monitor infants' physical and cognitive developmental milestones remotely. Experimental results from simulated lab tests show that computer vision can automate the process used by paediticians to monitor a baby's ability to follow objects visually with their eyes or move their head. Limited resources and access to health care in developing countries has reduced the frequency of postnatal care visits and subsequently increased infant mortality rates. In this way the proposed solution supports the global sustainable development goals of promiting healthcare for all.

Impact of bowel sounds on symptoms and colonoscopy outcomes

We utilized a modified stethoscope to capture digital recordings, amassing over 500 minutes of real patient data. Alongside these recordings, we diligently documented pertinent metadata from each patient, including specific symptoms, their Bristol scale ratings, and relevant medical antecedents. We employed machine learning algorithms with dual objectives: firstly, leveraging deep learning, we used a sequence of two Convolutional Neural Networks to craft a bowel sound recognition model that boasts an accuracy of 80%; and secondly, utilizing statistical models to evaluate the feature importances in predicting bowel sounds and symptoms. By engineering sound features, we were able to predict individual symptoms. Notably, our findings revealed that certain sound features possess p-values less than 1% when correlated with bowel symptoms. The overarching aim of this research is to unearth potential correlations between bowel sounds, manifested symptoms, and outcomes from colonoscopy examinations. Advancing research on the relationship between bowel sounds and bowel disorders holds the potential to revolutionize gastroenterology.

Improving Global Access to Maternal Healthcare using AI and Ultrasound

Every day, hundreds of women around the world die from preventable causes related to pregnancy and childbirth. The majority of these deaths occur in Africa, and most could have been avoided through interventions identified via ultrasound examination. AI can dramatically reduce the amount of training required to conduct maternal and fetal health ultrasound, potentially expanding global access to screening and improving health outcomes in underserved communities and countries. This talk will demonstrate the utility of AI-interpreted fetal ultrasound, even when using portable, low-cost probes and novice operators with no prior ultrasound experience. Results show that our on-device models are non-inferior, and in many cases, outperform sonographers on two critical diagnostic tasks: estimating fetal age and fetal malpresentation.

Day 2 - Opening Remarks

The African NLP Moonshot: Start your Rockets

Exploring the Frontier: The Challenges and Opportunities in Building Kenya's National Language Corpus

In the last two years, a team of researchers has taken the initiative to develop KenCorpus, an impressive open-source collection of textual and spoken data in three prominent Kenyan languages: Swahili, Dholuo, and Luhya. Furthermore, other individuals have taken the lead in gathering a Swahili dataset through Mozilla's Common Voice (MCV) platform, utilizing crowdsourcing. These datasets serve as fundamental resources for developers who aim to build applications like chatbots or automatic translation services. But how do you make use of such datasets as an aspiring NLP developer? How can such a repository further grow with and for communities?

Bridging Linguistic Frontiers: Machine Learning & NLP Innovations Empowering African Languages: Challenges, Progress, and Promising Futures

Whether expressed written, spoken, or signed, language is crucial for human communication and ensures understanding between people across regions. With the growing awareness and effort to include more low-resourced languages in NLP research, African languages have recently been a major subject of research in natural language processing. Accounting for more than 31% of living spoken languages, African languages are morphologically, culturally rich, diverse, and low-resourced. Covering a range of topics from (multilingual) machine translation, speech recognition, language modeling, named entity recognition, part-of-speech to datasets, and Lanfrica; in this presentation, I will share my journey into AfricaNLP research, challenges faced, progresses made, and future insights.

How AI and Virtual Reality Can Help Revolutionize Cultural Heritage in Africa

AI and Virtual Reality (VR) have the potential to revolutionize the preservation and dissemination of Africa's rich cultural heritage. AI can assist in the digitization and restoration of ancient artifacts, making them accessible to a global audience. VR technology enables immersive experiences, allowing individuals to explore historical sites and museums virtually. By combining AI's data analysis capabilities with VR's immersive power, we can create educational and engaging experiences, bringing African history, art, and culture to life for a diverse and global audience. This convergence of technology offers a unique opportunity to safeguard Africa's heritage and promote cultural appreciation worldwide.

Lunch Break

Creating beautiful art with Stable Diffusion

KerasCV and KerasNLP from Keras API are easy-to-use libraries for state-of-the-art computer vision and natural language processing. I will show developers how with just a few lines of code, they can employ the latest techniques and models for data augmentation, object detection, image and text generation, and text classification. I will demonstrate API integration with the broader TensorFlow ecosystem including TFLite. The workshop will also walk the developers through enabling diffusion models to generate images, explores some advanced uses of text to image models. It will show developers how to get started with generating images using KerasCV, the most optimized implementation of StableDiffusion available to date By attending this Keras API workshop, developers will gain practical knowledge and real-world examples that enable them to hit the ground running with KerasCV and KerasNLP. Workshop / Outcome It will show developers how to get started with generating images using KerasCV, the most optimized implementation of StableDiffusion available to date By attending this Keras API workshop, developers will gain practical knowledge and real-world examples that enable them to hit the ground running with KerasCV and KerasNLP.

African AI Pioneers: Charting the Course for Innovation and Impact

With Darlington Akogo; Vukosi Marivate; Alfred Ongere; Moses Kemibaro; Join us for a captivating panel discussion featuring African AI founders sharing their experiences, insights, and advice. Gain valuable perspectives on market trends, key considerations, challenges, and future outlooks in the AI landscape. This concise session promises to be both informative and inspiring, offering practical wisdom for aspiring entrepreneurs in Africa's burgeoning AI ecosystem.

Day 2 - Closing Remarks

Day 3 - Opening Remarks

Theory-to-Practice: Ensuring positive impact from ML solutions

With a growing adoption of machine learning (ML) solutions across different sectors, it is critical to ensure its positive societal impact. In this work, I will share key practices including the role of domain experts in the process and data valuation for potential systematic deviations to achieve trustworthy ML. Similarly, I reflect on our recent works across climate/sustainability and healthcare including automated identification and characterization of systematic deviations for various tasks, including data quality understanding, temporal drift, treatment effects analysis, and new class detection. Furthermore, I will also share how our generalized approach helps to evaluate capabilities of generative models in domain-agnostic and interpretable ways. I argue that similar data-centric analysis should also extend to traditional data sources beyond curated ML datasets.

AI for Biodiversity: Combatting Extinction Together

We are in the middle of the 6th extinction, losing the planet’s biodiversity at an unprecedented rate and scale. In many cases, we do not even have the basic numbers of what species we are losing and how many. New data collection technology, such as GPS, high-definition cameras, UAVs, genotyping, and crowdsourcing, are generating data about the living planet that are orders of magnitude richer than any previously collected. AI can turn these data into high resolution information source about living organisms, enabling scientific inquiry, conservation, and policy decisions. The talk will present a vision and examples of trustworthy AI for biodiversity, discussing opportunities and challenges.

Weather technologies without borders: Closing Africa’s infrastructure gap

Africa faces a critical weather data gap, hindering effective responses to extreme events. This talk explores how Google Research Africa is using machine learning techniques to close this gap. We’ll discuss new tools for forecasting precipitation in Africa and assessing damage, helping to reduce the human cost of climate-driven disasters like floods.

Interlude

Geospatial AI foundation models for Climate change discovery

Foundation models are artificial intelligence (AI) models that are pre-trained on large unlabeled datasets through self-supervision and then fine-tuned for different downstream tasks. There is increasing interest in the scientific community to investigate whether this approach can be successfully applied to domains beyond natural language processing and computer vision to effectively build generalist AI models. Here, we introduce Prithvi, a geospatial AI foundation model pre-trained on large source of multispectral satellite imagery from the NASA Harmonized Landsat-Sentinel 2 (HLS). Prithvi is a Temporal Vision Transformer that includes positional and temporal embeddings, which was trained on IBM Cloud Vela cluster (NVIDIA A100 GPUs) using a Masked Auto Encoder approach and Mean Absolute Error loss function for a total of 10k GPUs hours. Benchmarking downstream tasks such as flood mapping and burn scar identification, Prithvi could successfully be fine-tuned to produce state-of-the-art AI models for Earth observation tasks with the potential to achieve peak performance on test data quicker and with less training. As example use cases, we consider the applicability of geospatial AI foundation models to monitor reforestation activities in Kenya’s National Water Towers. The pre-trained model and fine-tuning workflows are available open-source on Hugging Face (https://huggingface.co/ibm-nasa-geospatial).

How WWF leverages AI for its work?

With Attila Steinegger; Confrey Alianji; Jonas van Duijvenbode

Sustainable water management

Despite abundant water resources, millions across Africa lack access to clean water due to inefficient management, infrastructure limitations, and uneven distribution. This presentation explores the critical challenges of water scarcity in Africa and examines how Artificial Intelligence (AI) emerges as a powerful tool to create a more sustainable future. We delve into key statistics illustrating the magnitude of the problem, highlighting factors like limited access to safe drinking water, reliance on unimproved sources, and the impact on health and development. We then explore existing solutions like infrastructure development, conservation programs, and decentralized options, discussing their limitations and the need for innovation. The presentation shines a light on the transformative potential of AI in water management. We showcase diverse applications, including AI-powered leak detection to reduce non-revenue water, smart irrigation platforms for optimizing agricultural water use, and flood and drought prediction models for improving preparedness. Taking a deep dive into AI-powered leak detection as a case study, we analyze its benefits, challenges, and sustainability considerations. We emphasize the importance of responsible implementation, capacity building, data privacy, and accessibility to ensure AI benefits all communities. The presentation concludes by urging stakeholders to embrace AI as a valuable tool for tackling water scarcity in Africa. By overcoming challenges and prioritizing ethical and sustainable use, AI can contribute to water security, improve public health, and boost economic development across the continent.

Lunch Break

Interlude: Rhythms of the continent

African Guarantee Fund: Data management for loans

Democratizing Deep Learning: The On-Device AI Revolution

Deep neural networks (DNNs) are revolutionizing computing, necessitating an integrated approach across the computing stack to optimize efficiency, especially for on-device deployment. In this talk, I will explore the frontier of DNN optimization, spanning algorithms, software, and hardware. We’ll start with efficient hardware-aware neural architecture search, demonstrating how tailoring DNN architectures to specific hardware can drastically enhance performance. I’ll then delve into the intricacies of DNN-hardware codesign, revealing how this synergy leads to cutting-edge hardware accelerator architectures. This talk aims to shed light on the pivotal role of codesign in unleashing the full potential of next-generation DNNs, paving the way for continued breakthroughs in on-device deep learning.

Advancing Africa's Climate Resilience and Agricultural Sustainability Through Environmental Data & AI

In most African countries, natural capital accounts for between 30 percent and 50 percent of total wealth. The continent is endowed with 65% of the world’s uncultivated fertile land and 30% of its mineral resources, yet it contributes only 3% to global GDP. In addition, frequent food and water scarcity still plague the continent despite having such enormous resources. One reason for this is the lack of reliable and trustworthy data, which has held back Africa’s development for decades by hampering business decisions and capital allocation, as well as making it difficult to measure the impact of climate change. In this session, we will explore how Amini.ai is harnessing the power of Geospatial data and Artificial Intelligence to address some of Africa's most pressing and urgent challenges, with a focus on adapting to climate change and enhancing food security across the continent.

The Role of AI In Shaping Of Industry And Economic Impact On Emerging Markets

Event Retrospective & Closing Discussions

The Future of Entrepreneurship: Leveraging Generative AI

Join me in exploring the exciting intersection of entrepreneurship and cutting-edge technology in my talk, ‘The Future of Entrepreneurship: Leveraging Generative AI.’ In this engaging session, I will delve into the transformative power of Generative AI and its profound impact on the startup landscape in Africa and globally.

No GPU No Problem!

As businesses gather massive amounts of data, running and deploying advanced machine learning models at scale is critical to gaining valuable insights. Baseten is an open source machine learning infrastructure platform that unlocks the power of data by enabling you to build, deploy, and run ML applications with models of any size and complexity. In this talk, I will give an overview of how Baseten can help businesses unlock the potential of their data using state-of-the-art machine learning techniques. Baseten supports transformer models with billions of parameters as well as small models and provides low-latency APIs to access them. You can train, tune, evaluate, and deploy models of any framework through a single simple interface. Baseten auto-scales infrastructure to maximise performance so you can go from prototype to production instantly. I will demo how Baseten powers data-driven applications with complex models by enabling you to query, retrain, and optimise large models on demand without the need for GPUs. You will learn how Baseten can help your business gain a competitive advantage by building advanced ML applications that tap into huge datasets inferring from a library of LLMs. An interactive Q&A session will follow, where you can ask questions about how to leverage Baseten for your data and models. Baseten provides an easy to use yet powerful platform for bringing complex data-fueled models to life. Whether you want to implement transformer models that achieve state-of-the-art results or build custom models tailored to your business needs, Baseten allows you to experiment, deploy, and scale ML models to unlock new possibilities with your data.

Privacy-Preserving Technologies for Africa

In an era where data is a valuable asset, the talk explores the critical role of privacy-preserving technologies in ensuring data sovereignty for African nations. Drawing insights from the latest advancements in federated learning, differentially private deep learning, and trajectory publication techniques, this session aims to delve into practical applications and challenges of these technologies in Africa. The discussion encompasses the implementation of privacy-preserving measures in diverse domains, from medical image analysis and the banking sphere to trajectory-based data, and other sensitive areas. We also highlight the stringent privacy policies adopted by institutions abroad as in Canada as a benchmark for safeguarding sensitive information. Moreover, the talk will shed light on recent developments in Africa, emphasizing the importance of AI safeguards to protect online privacy. Attendees will gain practical knowledge, including the use of free digital tools, to preserve their data rights in an evolving digital landscape.

GeoAI in Africa: Harnessing Satellite Data for Climate Resilience

With Akram Zaytar; Gilles Q. Hacheme; This workshop is aimed at demonstrating the utility of geospatial machine learning in addressing large-scale food security issues in Africa. Participants will learn how to construct an end-to-end ML pipeline for mapping crops in a region of interest using satellite imagery. The workshop will cover key concepts of geospatial machine learning such as data acquisition, spatial analysis, data engineering, machine learning, and inference. Focusing on agriculture to mitigate the effects of climate change in Africa. Workshop / Outcome By the end of the workshop, participants will have a foundational understanding of geospatial machine learning applications in agriculture. They will be equipped with practical skills to utilize satellite imagery for automated spatial mapping and have insights into forming collaborative networks to address climate-related challenges in Africa. Workshop / Difficulty Beginner level Workshop / Prerequisites - Basic to intermediate knowledge in Python and machine learning. - Familiarity with satellite imagery and geospatial data is beneficial but not mandatory. - Installation of necessary software and tools as detailed in the workshop's GitHub repository README file. - Participants are encouraged to install and set up the required tools prior to the workshop for a more efficient hands-on session.

ASR for informal retailers

Speech recognition technology faces significant challenges in real-world use cases, particularly in African contexts such as informal retail. Code-switching, the seamless transition between languages or dialects within conversations, poses a major obstacle to accurate transcription. Additionally, limited access to diverse and labeled speech datasets, compounded by factors like background noise and accent variability, hinders model training and performance. Innovative approaches leveraging techniques such as data augmentation and crowdsourced dataset curation are essential to address these challenges and enable the widespread adoption of speech recognition technology in Africa.

Machine learning and spectral matching techniques improve landscape-scale crop type classification

Accurate crop-type mapping is essential for improving food security, promoting sustainable agricultural practices, and shaping agricultural policies. However, obtaining the requisite landscape scale data for crop type mapping is resource-intensive, particularly in resource-limited countries and expansive geographical areas. This study presents an innovative spatial explicit methodological framework for crop-type mapping in the heterogeneous landscape of the Western Cape Province, South Africa. The approach integrates unsupervised learning and spectral matching techniques. The Learning Vector Quantization (LVQ) algorithm was employed to cluster Sentinel-2 Multispectral Instrument (MSI) imagery acquired between April 2021 and March 2022 into distinct monthly crop clusters. Subsequently, spectral matching techniques were applied to assign accurate crop types to these clusters. The LVQ algorithm revealed the dynamic spatiotemporal variations in cluster counts that mirror the region’s agricultural diversity and climate patterns. The spectral matching exhibited varying effectiveness in associating crops with clusters. May (Overall Accuracy (OA) = 0.84, p = 0.01) had the highest accuracy, due to differing crop growth stages and canopy cover conditions during this period. Despite limitations, including the medium spectral resolution of Sentinel-2 MSI, these results advance cost-effective agricultural monitoring and land management practices. The results establish a robust baseline for comprehensive crop-type mapping contributing to informed agricultural decision-making. In addition, these results underscore the transformative potential of remote sensing and machine learning in enhancing crop management precision and sustainability. The implications of this study extend beyond food security and address temporal gaps in conventional crop surveys.

Intelligent Document Processing - Your First Step Towards Digital Transformation

OCR technology has a direct and transformative impact on numerous industries, including healthcare, finance, retail, governance, and more. The combination of intelligent document processing (IDP) and OCR presents a powerful synergy that can significantly expedite the digital transformation process in the African market. Many businesses in Africa still rely on archived papers or stored PDFs, containing a wealth of valuable information. Leveraging these data reservoirs can provide organizations with deeper insights, streamline processes, and automate tasks with remarkable efficiency. IDP platforms have the potential to convert these archives and documents into dynamic, indexed data. They can extract valuable information, potentially fueling the development of online applications, such as banking and insurance apps. Automating manual data entry not only boosts productivity but also reduces the risk of human errors, saving organizations time and instilling an AI-first mindset.In my upcoming presentation, I will delve into real-life case studies where industries have harnessed the power of Intelligent Document Processing (IDP) and Optical Character Recognition (OCR) to transition from manual processes to seamless automation. Furthermore, I will provide insights through benchmarking, sharing relevant statistics and highlighting the distinctions between the African market and counterparts in Europe, the Pacific, and North America. I will cover key players in this technology landscape, the top industries benefitting from these solutions, the current state of technological advancements, the growth in the number of digitized documents over the years, and the return on investment (ROI) achieved through these transformative technologies.