#ICML2023 invited talk: Shakir Mohamed on ML with social purpose

by Lucy Smith

29 August 2023

Machine learning for earth systems – modelling weather and climate

Modelling of weather and climate can have a big impact on society, with such models often providing the basis for decisions taken by policy makers. These could be short-term decisions (such as the evacuation of an area due to a sudden flood risk due to heavy rain) or policy decisions based on long-term climate modelling.

To model these processes, researchers use Earth system models (ESMs), which take into account a range of physical, chemical and biological components. At their core are atmospheric and ocean processes, and they also include representations of the global carbon cycle, dynamic vegetation, atmospheric chemistry, and continental ice sheets. ESMs simulate how different components change over time.

Shakir and his team have been applying machine learning to ESMs with a focus on nowcasting (high resolution predictions 1-2 hours ahead) and medium-range forecasting (up to 10 days ahead). With nowcasting, of special interest for meteorologists is heavy rainfall prediction, as it is the weather phenomenon most likely to affect lives and property with little warning. Therefore, good nowcasting models should excel at predicting heavy rain events.

Building a successful model depends on having access to data from ground station radar networks that measure the amount of moisture in the atmosphere. The UK, for example, has radar that covers 99% of the country with 1km grid resolution and a data feed that updates every five minutes. The data is presented in video form with each pixel representing the accumulated moisture level in units of mm per hour.

Traditionally, ESMs have been based on physical equations, and numerical and deterministic models. However, to construct their model, Shakir and colleagues used a generative adversarial network (GAN) approach. Working with meteorological experts they tested their model on the notoriously difficult to predict heavy rain events in East Scotland. The team also used the GAN approach to construct a base model for medium range forecasting. The task was to predict atmospheric state at six hour intervals at high resolution for the following 10 days. Their models showed state-of-the-art performance compared to existing deterministic models. Shakir sees these generative models not as a replacement for, but rather a support for, existing numerical and physics-based models.

A sociotechnical approach

To introduce his second theme, sociotechnical AI, Shakir used the field of forecasting to illustrate that, often, AI models that were intended to improve lives end up having the opposite effect. For example, in Brazil and Zimbabwe, better rainfall forecasts which were developed to help farmers ended up being used by bank managers to deny credit to farmers who would otherwise have received it. In Peru, better forecasting of El Niño and the prospect of a weak season gave fishing companies an incentive to accelerate the layoff of workers. Unfortunately, better models have led to greater harm and vulnerability.

Shakir emphasised that no technical system exists independently of the social world. This interaction between technology and social considerations is referred to as a sociotechnical approach. Shakir introduced the concept of a “sociotechnical stack” which includes four different layers that researchers should consider, namely:

Community and cooperation
Governance and accountability
Deployment and commercialisation
Research and innovation

Essentially, technical and engineering work should account for a wider set of considerations. This opens up new areas of research, and Shakir urged the community to contribute, intervene, and act with regards to the different layers of this stack.

Shakir is an advocate for participatory approaches to AI design, and by this he means including people in the design of methods, being comfortable with disagreement, and being open to changing working practices and even topics. He noted that the field is making progress in this respect, but community participation is still minimal.

Towards global AI for all

This brought us to the final strand of the talk, considering how we can make AI truly global, and truly for everyone. To achieve this, technology that supports society and culture, rather then becoming an instrument of cultural oppression and colonisation, is essential. Shakir believes that there is power in strengthening varied forms of political community, where intercultural dialogue is at the forefront.

Shakir has experience of putting this into practice. About seven years ago, he was a founder of Deep Learning Indaba, which has the mission of strengthening machine learning across the continent of Africa. Over the years they have built new communities, fostered leadership, and recognised excellence in the development and use of AI across Africa. AI is becoming more global because of the efforts of groups such as Deep Learning Indaba.

A call to action

In closing, Shakir encouraged the audience to embed three actions within their research. Namely to 1) infuse their work with social purpose, 2) develop their view of the sociotechnical stack and intervene where possible to support a richer participation in AI, and 3) support grassroots efforts to achieve truly global AI.