Leveraging AI

The African continent has seen the growth of a number of high-impact innovations across several domains that are leveraging emerging technologies to solve pressing problems, and Artificial Intelligence (AI) is a major driver for these innovations.

In Uganda at the Makerere AI lab,1 we have been involved in several AI for social good innovation projects. For example, one project is working on using mobile microscopy to diagnose malaria and tuberculosis using a mobile phone.2 We have also built AI models and deployed these on mobile phones for the detection of viral diseases in cassava and beans crops. Our applications can be used to monitor disease spread across the country in near real-time. Recently, we also launched Luganda on the Common Voice platform as a way of crowdsourcing a diverse voice dataset that is important for building speech recognition models.3 Consequently, we are using this data to train Deepspeech models for automatic speech-to-text for Luganda.

Africa has a unique opportunity to harness the growth of AI across different fields and tackle some of the economic challenges the continent is facing. However, for such success to be achieved, several aspects of AI have to be taken into consideration to drive the full potential of AI for the continent, in particular the data-to-impact pipeline for creating AI solutions and systems.

This data-to-impact pipeline involves many steps, including the AI problem formulation, data collection and curation, model development, field tests and deployment, feedback and maintenance and impact and sustainability. We will not attempt to address all the steps on the pipeline but will focus instead on the most critical issues to harness innovation within the AI ecosystem.

Problem Identification. Identification of the initial ‘problem’ to be solved must be rooted within the community for which the AI solution is being developed and it should be a "real-world problem."
As an example of a real-world problem, one of the projects undertaken by the Makerere AI lab on agriculture highlights that the problem for many farmers is a lack of knowledge about crop diseases and the need for an agricultural expert to provide a quick diagnosis. Given the limited number of agricultural experts, often by the time an expert is able to visit the smallholder farm, the crop diseases have spread. In this case, problem identification involves consultation in a participatory and interdisciplinary approach with relevant stakeholders like agricultural experts, extension workers, researchers, smallholder farmers. Data scientists talk to these domain experts to understand the challenges they face, any constraints present in the domain, the environment and systems in which they operate and the kind of data that they possess. Each stakeholder has tasks in the data-to-impact pipeline and all interact together to formulate the problem, create the machine learning datasets and recommender models to use the models to solve an issue affecting them in their community.

Data collection and sharing: According to UNESCO (2021)4 one of the major issues with building AI models in Africa is access to AI training data, and the lack of localised training data overall, which affect the use of AI systems. Training data needs to be diverse, localised and contextualised to eliminate bias that might result from the under-representation of some populations - for example, the lack of minority language datasets that represent the languages spoken by diverse populations on the continent. AI training datasets should also be available and shareable with the wider AI community, making them central to AI development research and innovation.5
Data sharing should be done from the “bottom-up” by the dataset creators taking into consideration several aspects around where the data comes from and the cultures, lives, communities, context that the data represents.6 This also involves thinking about ethical considerations and implications with the data collection and sharing process. For example, the Open for Good initiative7 provides a platform for the coordination and exchange of good practices around building and sharing localised AI training data.
Context matters: Understanding the context when collecting and sharing data, and building and deploying AI models is critical.8 For example, when building an AI model for breast cancer diagnosis, contextual factors such as available treatment options, unique genetic markers, or family history should be taken into consideration. AI solutions that are obtained with technology that is locally developed within the continent is important for context-driven models.9 All relevant stakeholders need to be included at an early stage of the design of local and effective AI solutions.
Ethical considerations in education and research: While AI holds promise, taking ethical concerns around algorithmic bias into consideration, or incorporating a feminist AI perspective could help support trust.
From an education perspective, this means developing training programmes for the inclusion of AI ethics in education at different levels through the computing and engineering curricula.10
Implementation guidelines can also be put in place to leverage working use cases to translate AI ethical principles into practice.11 The hope is that across the AI solutions being piloted that they can adapt feminist practices around the solutions to support the data-to-impact pipeline.

Artificial Intelligence in Africa presents unique opportunities to put the continent at the forefront of AI innovation. But the focus when building AI systems should not be solely on the outcome but also on the unique journey that leads to the outcome for each of the AI solutions and systems.

References

Abebe, Rediet, Kehinde Aruleba, Abeba Birhane, Sara Kingsley, George Obaido, Sekou L. Remy, and Swathi Sadagopan. ‘Narratives and Counternarratives on Data Sharing in Africa’. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 3 March 2021, 329–41. https://doi.org/10.1145/3442188.3445897.

‘Automated Laboratory Diagnostics - AI Research Group, Makerere University’. Accessed 19 May 2021. https://air.ug/microscopy/.

Bangani, Siviwe, and Mathew Moyo. ‘Data Sharing Practices among Researchers at South African Universities’. Data Science Journal 18, no. 1 (3 July 2019): 28. https://doi.org/10.5334/dsj-2019-028.

Birhane, Abeba. ‘Algorithmic Colonization of Africa’. SCRIPT-Ed 17 (6 August 2020): 389–409. https://doi.org/10.2966/scrip.170220.389.

Borenstein, Jason, and Ayanna Howard. ‘Emerging Challenges in AI and the Need for AI Ethics Education’. AI and Ethics 1, no. 1 (1 February 2021): 61–65. https://doi.org/10.1007/s43681-020-00002-7.

‘Common Voice by Mozilla’. Accessed 19 May 2021. https://commonvoice.mozilla.org/.

‘Makerere - AI-LAB’. Accessed 19 May 2021. https://air.ug/.

‘Open for Good Alliance | AI Training Data for Developers’. Accessed 19 May 2021. https://www.openforgood.info/.

Sibal, Prateek, and Bhanu Neupane. ‘Artificial Intelligence Needs Assessment Survey in Africa | IRCAI’. Accessed 19 May 2021. https://ircai.org/project/unesco-ai-needs-assessment/.