AMLD Africa 2022

Opening Ceremony

[Invited Keynote] Connecting the Dots: Graph-ML and Beyond

Graph data models are increasingly being leveraged across different critical enterprise workloads. The graph model easily captures data relationships that most data models fail to capture, along with enabling data aggregation from various heterogeneous data assets. This keynote presents Oracle's take on Graph and ML advances along with our journey through industry macro-trends and micro-trends. We also provide practical use cases and business problems that we are solving in healthcare and cyber-security using graph data and ML models and discuss our vision for future enterprise grade data analysis platforms

[Invited Keynote] - Reimagining medicine at Novartis with the help of AI

We are witnessing a rapid evolution of the healthcare ecosystem boosted by new technologies to better quantify health and bring data driven insights into clinical and pharmaceutical decision making. We will outline the significant challenges associated with conceiving and brining a new drug to market, and highlight the potential that machine learning offers to augment this process via a growing amount and diversity of health data. In particular, personalized and precision medicine are becoming a must for better care. We will dive into specific examples on how machine learning is tailoring drug development via recommended patient characteristics and subgroups, as well as the challenges of working with health data. And lastly, how fresh talents can take their first steps towards launching a career in this space.

Late fusion of multi-magnification histopathological breast cancer images using deep learning

Breast Cancer is the most frequently diagnosed malignancy and the leading cause of cancer death among women worldwide, the main strategy to reduce mortality is early detection and treatment. Histopathology biopsy imaging is currently the clinically practiced gold standard for diagnosing breast cancer since it captures the effect of the disease on the tissues with a comprehensive view. Pathologists examine the images at various magnification factors to identify the type of the tumor, because if just one magnification is taken into account, the decision may not be accurate, especially if there are variances in the patient’s score at each magnification level. Thus, machine learning models must process the image at various magnification factors for an accurate classification of breast cancer tumors. This work explores the performance of transfer learning and late fusion to construct a multi-modality ensemble that fuses different magnification-specific deep learning models in order to predict the class of the breast cancer histopathological slides.

[Invited Keynote] Panning for insights in medicine and beyond: New frontiers in machine learning interpretability

In this keynote, I describe an extensive new framework for ML interpretability which enables us to turn black-box machine learning methods into white boxes. This framework allows us to unravel underlying governing equations from data, enabling scientists to make new discoveries. Finally, I will introduce our extensive github for ML interpretability: https://github.com/vanderschaarlab/Interpretability

Using Big Data to drive Health Outcomes

With the widespread adoption of mobile phones, low-cost computing and internet access, significant progress has been made in the digitisation of healthcare systems in emerging markets. More data is being created every day, and the challenge has now become: how can we make use of this data to drive positive outcomes?

Leveraging data from engaged users to improve maternal and prenatal health using medical grade wearables

In this talk we will show how our data driven approach helps to shed light on poorly understood links between physiological changes and pregnancy outcomes. We will discuss opportunities arising from crowdsourcing clinical research, by providing consumers with clinical grade tools and data, and analyzing such data at a scale beyond what is possible in regular clinical studies. We will take specific examples from Wekebere's research on maternal health, fetal wellbeing and labor detection.

[Invited Keynote] Machine Learning for Scientific Discovery

Machine learning research is expanding its reach, beyond the traditional realm of the tech industry and into the activities of other scientists, opening the door to truly transformative advances in these disciplines. In this lecture I will focus on two aspects, modeling and experimental design, that are intertwined in the theory-experiment-analysis active learning loop that constitutes a core element of the scientific methodology. Computers will be necessary to go beyond the currently purely manual research loop and take advantage of high-throughput experimental setups and large-scale experimental datasets. I will discuss methods related to active learning, reinforcement learning, generative modeling, Bayesian ML, amortized variational learning and causal discovery. I will discuss the notion of epistemic uncertainty and how to estimate it. I will motivate generative policies that can sample a diverse set of candidate solutions to a problem, be it for proposing new experiments or causal hypotheses. Finally, I will describe current research to help us with these questions based on a new deep learning probabilistic framework called GFlowNets and how we plan to apply these in areas of great societal need like the unmet challenge of antimicrobial resistance or the discovery of new materials to help fight climate change.

Antipatterns in machine learning in healthcare

This talk aims to discuss issues and common pitfalls during the lifecycle of research projects in the field of machine learning in healthcare. There is a huge gap in the promise and adoption of ML systems in the clinical domain and this talk aims to point out some of the reasons, problems and solutions surrounding this very critical space.

Enzeptional: enzyme optimization via a generative language modeling-based evolutionary algorithm

Enzymes present remarkable characteristics to enable more sustainable reactions. Our capacity to design effective enzymes in silico has not yet been brought up to speed for industrial research. By viewing enzyme optimization problem as a language modelling task, we can adapt active sites to perform new reactions.

Land boundary delineation using u-nets

Land delineation is a particular example of a segmentation problem. In this project we used U-Nets, a type of CNN, to learn the boundaries of parcels of lands from satellite images. The U-Nets model proved very efficient for the task.

Machine Learning approaches for cross-continent transfer of Remote Sensing models for Africa

Available Remote Sensing datasets are unequally distributed: The Global North has much more labeled data available compared to the Global South. In the project “PreTrainAppEO” we are investigating possibilities of leveraging training data from data-rich regions for applications in data-sparse ones, using recent transfer- and meta-learning approaches.

Improved water resource management framework using remote Sensing data and machine learning methods

The introduction presents the context of water scarcity in Morocco and the problem of our thesis. The methodology shows water resources management using rain gauges and remote sensing data as an input in the model. the typical parameter and variables for the model. and finally results and discussion

[Invited Keynote]

[Invited Keynote] Building Trustworthy Machine Learning Systems

Can Natural Language Modeling be democratized by moving from the classical MLM pre-training paradigm to the Multi-task pre-training one?

Up to now, Masked Language Model (MLM) framework has been widely adopted for self-supervised language pre-training using large-scale general corpus. Moreover, due to the complexity of natural language, we are witnessing a race for increasingly massive models and this cannot be sustainable. In contrast, there are works arguing the benefit of diversifying tasks in the pre-training phase and other works focusing on efficient ways to positively share knowledge between tasks. This efficient diversification in tasks would potentially give birth to new pre-training paradigms enabling richer language representation and not requiring as much parameters as the classical MLM pre-training paradigm. In this talk, we propose to review these two sets of works namely multi-task pre-training and effective knowledge sharing between tasks, and show some of their implications in the case of Arabic language and its dialects.

Transfer Learning from Modern Standard Arabic to Dialectal Arabic

With the widespread adoption of natural language processing (NLP) systems through chatbots, writing assistance, and translation... Under-represented languages are still far from having such achievements. This was due to the fact that large amount of data is needed to build such systems, which is not accessible for all languages. Therefore, the challenge has now become: how can we leverage the new NLP achievements to low-resource languages? Abdellah will discuss how to use machine learning to build NLP Systems for low-resource languages, he will present this for the case of the Arabic language and its varieties. These approaches show to achieve high performance for low-resource languages using Zero-Labeled data.

Building language models, one story at a time

One-third of the world's languages are spoken in Africa, but less than 1% of African languages are represented online. This is significant because the language you speak, write or sign shapes your online experience. In a Google Research collaboration in Mali, we worked with griots to build a language model for Bambara, one of West Africa's most widely spoken languages. Griots are historians, storytellers, and the real keepers of the Bambara collective memory, passing their knowledge only through storytelling. By recording griots narrating generational stories, we were able to build a model for understanding and translating Bambara speech. We release the findings for the research community and everyone to benefit.

Speech Improvement Framework using LSTM Neural Network and Dynamic Time Warping to improve Kenyatta National Hospital records retrieval

In healthcare, speech recognition is a dominant application of deep learning algorithms. The current systems are more involved in words that are proposed correctly and available in the dictionary, however, algorithms to detect and correct errors are still a challenge. This study will seek to develop a framework based on Long Short-Term Memory (LSTM) neural network and Dynamic Time Warping (DTW) algorithm which also enhances performance on speech signals. The developed framework will also address the issue of accuracy as well as word error rate which is a measure of the performance of speech recognition systems. LSTM neural network is known to have the ability to counter the vanishing gradient problem experienced by recurrent neural networks (RNN) whereas Dynamic time warping is effective in calculating the resemblance between two-time series. This study will adopt the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology. The developed algorithm will be tested and validated using health records thus giving better performance and increased accuracy.

[Invited Keynote] If it isn’t diverse, it isn’t ethical

Diversity in AI is essential to ensure that the technology we build is fair and just. We must take a hard look at who creates, designs and implements our products and consider how these profucts might impact our society at large. So what does this mean for the future of AI? If we want artificial intelligence to work well, it needs to reflect the societies it serves. It means using artificial intelligence to help us solve the problems we face as a society, rather than just making our lives easier.

[Invited Keynote] Artificial Intelligence in the European Union

[Invited Keynote] AI for industrial and chemical processes

Feature Reduction on FT-NIR Spectroscopy to detect Bruised Apples

Spectroscopy data are useful for modelling biological systems. However, using a wide spectrum of wavelengths is not practical in a production setting. Variable selection methods are one efficient way to obtain an optimal model and were the aim of this work. Near-infrared spectral data in the range of 800 – 2500 nm were used to classify bruise damage. Six machine learning classification algorithms were employed, and two variable selection methods were used to determine the most relevant wavelengths for the problem of distinguishing between bruised and non-bruised apples. The selected wavelengths clustered around 900 nm, 1200nm and 1900 nm. The best results were achieved using linear regression and a support vector machine based on up to 40 wavelengths.

A mobile application to empower smallholder farmers and build bridges between farming & data

SOWIT is an Agtech startup living at the crossroads of remote sensing, AI and agronomy, which goal is to build products fitted to serve the African smallholder farmers, through intense field iterations and user centric design.

Optimizing crops and application to Kenya and other countries in Africa

We combine (i) Geopy’s geographical management tools (ii) the models of crops and crop varieties in PCSE (iii) the optimization capabilities of Nevergrad (iv) NASA’s climate data We generate maps with, at each location: (i) a recommended crop and a recommended variety. (ii) irrigation schedules. Parameters include year (possibly average over several years) and irrigation possibilities, so that we can output results per decade so that climate change is visible and take into account yearly weather forecasts.

Machine learning for crop disease detection a case study of cassava crop leaves

Cassava is one of Africa’s most important crops found in tropical regions worldwide. As important as cassava is, concerns are raised towards its diseases. In this paper, the aim is to identify and classify images of cassava leaves into healthy and unhealthy (disease) categories using machine learning. The categories include; four(4) disease categories - Cassava Bacterial Blight (CBB), Cassava Brown Streak Disease (CBSD), Cassava Green Mite (CGM) and Cassava Mosaic Disease (CMD), and one (1) health category – Healthy. At the end of this implementation, we were able to train a model that classifies images of the given cassava leaves into one of the five(5) categories.

[Invited Keynote] From Big Data to Better Data for Trustworthy AI

[Invited Keynote] Graph ML - The Next Level of Machine Learning

This talk focuses on why Graphs have become one of the biggest trends in Machine Learning. Graph Machine Learning based on Graph Analytic Algorithms is driving significant improvements in Fraud/Anomaly Detection, Ranking (Page Rank), Recommendation Engines (collaborative filtering), text summarization, and other NLP tasks. We will cover Graph Analytic Algorithms, their applications, and the more novel–but equally exciting–field of Graph Machine Learning, including topics such as Graph Embeddings, Graph Neural Networks, and applications of Graph Machine Learning. At the end of this session, you will better understand Graph Machine Learning techniques, when to use them, and when to consider other methods

Learning on Graphs with Explainable Graph Neural Networks

Graph ML provides powerful techniques to address the most difficult ML problems in which the data is connected and can form complex patterns. In this talk we will present Oracle's take on Graph ML advances, and how the different techniques can be applied in several domains.

Building with NVIDIA Omniverse: Create Tools for the Metaverse

The Metaverse is the 3D Evolution of the Internet. 3D virtual worlds and workflows, whether in design or simulation, are now an essential component of every industry, and will require a new set of tools: NVIDIA Omniverse is a scalable platform to build and operate Metaverse applications. In this session we will explore how you can align Omniverse to your business model and build the tools for the virtual worlds of today and tomorrow.

[Invited Keynote] How are organizations really using AI? A tour of use cases.

New developments in AI/ML are almost daily, with ground-breaking innovations like GPT-3 or Dall-E2 being announced at regular intervals. But how are organizations worldwide really adopting AI/ML to transform or optimize their business? This session will be a tour of some of the many cases Microsoft Azure AI customers are deploying to transform their organizations, from tabular to forecasting, NLP and Computer Vision.

Federated Learning with Azure Machine Learning

In this demo rich session about Federated Learning, we cover the following items: 1. Need to protect privacy and intellectual property in Machine Learning use cases 2. Understanding the Federated Learning Concept 3. Demo: Deploying and running a global Federated Learning setup on Azure ML

Use AutoML to create high-quality models

Automatic machine learning, known as AutoML, removes the tedious, iterative, and time-consuming work across the machine learning (ML) workflow from data acquisition to model operationalization, so you can spend less time on low level details and more time on using ML to improve business outcomes. In this session you will learn about the AutoML tools that ake care of sourcing and preparing data, engineering features, training and tuning models, deploying models, and ongoing model monitoring and updating.

Closing Ceremony

An Introduction to PyGlove & Neural Architecture Search

PyGlove is a recently open-sourced library by Google that allows programmatically changing code at runtime. This "makes meta-programs much easier to write" to quote the PyGlove development team. Which makes it easy and quick building of powerful AutoML/Neural Architecture Search algorithms. We will explore Neural Architecture Search using PyGlove as our library of implementation/execution. Delivered by: Ali Hussein, Daisy Khaabi,Mazen Mobtassem & Omar Abo El Fotoh, Participants will have an understanding of Neural Architecture Search and how to implement a NAS algorithm using PyGlove.

Introduction to Data-Centric AI

Data-Centric AI is an approach to building machine learning models by focusing more on data instead of code. In this workshop, we are going to learn how to leverage data-centric AI techniques to build good models. The outcome of this workshop is as follows: - Understand the importance of quality data on model performance - Learn techniques for improving data quality - Learn how to use tools such as DVC and IPython Widgets. The participants should have knowledge of building models using any deep learning framework. Preferably they should also have a basic understanding of computer vision.

Building an end to end ML pipeline

Data Engineering On AWS

Building Transformer-Based Natural Language Processing Applications

Applications for natural language processing (NLP) have exploded in the past decade. With the proliferation of AI assistants and organizations infusing their businesses with more interactive human-machine experiences, understanding how NLP techniques can be used to manipulate, analyze, and generate text-based data is essential. Modern techniques can capture the nuance, context, and sophistication of language, just as humans do. And when designed correctly, developers can use these techniques to build powerful NLP applications that provide natural and seamless human-computer interactions within chatbots, AI voice agents, and more. Deep learning models have gained widespread popularity for NLP because of their ability to accurately generalize over a range of contexts and languages. Transformer-based models, such as Bidirectional Encoder Representations from Transformers (BERT), have revolutionized NLP by offering accuracy comparable to human baselines on benchmarks like SQuAD for question-answer, entity recognition, intent recognition, sentiment analysis, and more. In this workshop, you’ll learn how to use Transformer-based natural language processing models for text classification tasks, such as categorizing documents. You’ll also learn how to leverage Transformer-based models for named-entity recognition (NER) tasks and how to analyze various model features, constraints, and characteristics to determine which model is best suited for a particular use case based on metrics, domain specificity, and available resources. Learning Objectives By participating in this workshop, you’ll: Understand how text embeddings have rapidly evolved in NLP tasks such as Word2Vec, recurrent neural network (RNN)-based embeddings, and Transformers See how Transformer architecture features, especially self-attention, are used to create language models without RNNs Use self-supervision to improve the Transformer architecture in BERT, Megatron, and other variants for superior NLP results Leverage pre-trained, modern NLP models to solve multiple tasks such as text classification, NER, and question answering Manage inference challenges and deploy refined models for live applications

Basics of Reinforcement Learning for industrial learning

We first introduce the important concepts of RL and Contextual Bandits, followed by their applications in industry, including a technical case-study blog where we applied it to boost user engagement for Digitec Galaxus's newsletter. In the second half, you will do a hands-on lab applying these concepts to make a small recommender system using bandits and RL on Google Cloud, and optionally test your knowledge with a quiz on the course's page on Coursera.

A guide on Blockchain Data Analysis and Model deployment

This workshop will go in-depth into how to get blockchain data from apis, visualize and analyze the sample data, train a machine learning model on sample data, and deploy the trained and tested model. After this session, participants will have learned how to get blockchain data from explorers(etherscan, solscan), NFT marketplace(opensea and looksrare) and exchanges(finance), data visualization, and analysis tips and how to train sample blockchain data and deploy the trained and tested model.

Time Series Applications. Participants will come away with some theoretical knowledge of time series data and techniques to model time series data, as well as knowledge of how to implement these techniques in Python. It is hoped that the Hackathon part will equip participants with the skills to solve real-world problems in their home countries that involve time series data--tracking climate change to predict future impacts, understanding market and financial instability from events such as the COVID pandemic, testing healthcare interventions over time on reduction of mortality, predicting future food prices in local markets... These can be the start of research papers, government grants, or businesses, and we're happy to help participants build out ideas after the workshop or connect them with those who can help build out the ideas in their home countries.

The first half of the workshop will focus on common time series algorithms from a theoretical and applied perspective, including ARIMA/singular spectrum analysis/recurrent neural networks for prediction, topological/geometric approaches for change-point detection, and longitudinal statistical models for statistical testing of factors influencing time series values. The applications will involve Python code and datasets such as the Moroccan stock market, food prices in Burkina Faso, health indicators in Gabon, and climate change in Sudan, among others. Jupyter notebooks and datasets will be provided ahead of time for students to follow along during this half of the workshop, and we'll be making the materials from the talk available to anyone who isn't able to attend but would like to learn. The second half of the workshop will involve a Hackathon using either the Humanitarian Exchange and Zindi open-source datasets from our examples or whatever data participants want to bring. Focus will be on applications that solve problems and could be scaled into papers or businesses. Build teams. Work independently. Come up with a solution in the workshop, or build out a longer-term project!