Programme

If you’ve ever taken a statistics class, I have little doubt you’ve heard someone say “correlation does not imply causation”. This trope seems to imply that we can never state that A causes B by analysis of data alone. This talk will take a deeper look at if this is always the case by introducing the field of causal inference, with a special focus on (graphical) causal modelling. Attendees will leave having gained knowledge of how assumptions and data come together to enable drawing causal conclusions from data. More importantly, they will develop the basic intuition and fundamental concepts of causal modelling in the context of machine learning.

Array algebra refers to the operations that we can perform on rectangular arrays (“tensors”) of numbers, which are the fundamental data structures used in deep learning, as well as more generally in data science and data-oriented programming. This talk, which pairs with the array algebra practical, will give a whirlwind tour of what arrays are, and help to organize the operations we typically perform on them. We’ll look at the natural sequence of n-arrays where scalars / numbers = 0-arrays, vectors / tuples = 1-arrays, matrices = 2-arrays, etc. We’ll review some concrete examples of different n-arrays, and what the axes of these arrays mean. We’ll triangulate among several different viewpoints of n-arrays, to gain a better intuition of the core array operations: generalized dot products, aggregation, reshaping, transposing, slicing, stacking, mapping, etc. We’ll also try to equip you with helpful ways of thinking about higher-order arrays and their axes, one of the more challenging aspects of coding with complex neural network architectures.

Many problems in machine learning require us to compute the sum or integral of a distribution. Examples of this include calculating the expected value of a random variable or computing the normalising factor in Bayes’ theorem. While there are certain cases where an analytic solution can be computed, most often an exact solution is intractable and so must be approximated.  In this talk, we cover Monte Carlo methods, where samples drawn from distribution can be used to numerically approximate the quantity of interest. We describe rejection and importance sampling, which can be used even when the target distribution is not easy to sample from. However, these approaches do not work well when applied to huge models or datasets, as is commonly the case in machine learning. We, therefore, conclude with Markov Chain Monte Carlo – an algorithm that can be used to efficiently sample from high-dimensional distributions.

• Explain it to your CEO
• Train a fellow machine learning engineer
• Maintain a conversation with a seasoned transformer researcher

We will also demonstrate how to test if they suit your problem quickly and, if they do, how to structure your problem and data to get them working.

Ever-increasing qualities and quantities of data are routinely collected concerning all aspects of our life.  Developments in wearable sensors, smart home technologies, and the Internet of Things provide the industry with opportunities to improve its services. However, analysing time-series data collected in different applications poses several challenges as the data can have substantial artefacts, might be incomplete, and contain high variations. This tutorial will introduce several machine learning techniques to tackle these issues and provide robust solutions to address such challenges. Various real-world examples especially on dealing with time series data in healthcare settings will also be provided.

Lacuna Fund: Agricultural Datasets and NLP Datasets

1. Eyes on the Ground: Smartphone Images for Crop Monitoring.
2. HornMT – Machine Translation Benchmark Dataset for Languages in the Horn of Africa
3. NaijaSenti: A Nigerian Twitter Sentiment Corpus for Multilingual Sentiment Analysis
4. Kencorpus: Kenyan Languages Corpus for NLP and Machine Learning

Zindi Presentations

1. Turtle recall 3rd place solution
2. East African Field Centers

Website: https://wscv-indaba.github.io/

https://wscv-indaba.github.io/