Trends in Frontier Tech: Machine learning in South Asia
During our sixth call for Frontier Technologies pilots we received 23 applications for innovations which contained an artificial intelligence component. Of these applications, 13 sought to test machine learning innovations in South Asia.
Machine learning refers to a specific branch of artificial intelligence which is concerned with using data and algorithms to imitate the way humans learn and gradually improve the accuracy of the solution in conducting a task. The applications identified a range of use cases for machine learning in South Asia, including in relation to agriculture, environmental management, EdTech and supply chain management.
These applications may to some extent be indicative of a high level of interest in machine learning within South Asia. Moreover, they indicated how catalytic machine learning solutions can be in undertaking processes and delivering insights far more quickly than any human could. Among those applications we received to use machine learning in South Asia there were two key sector specific trends. These are highlighted below.
Machine Learning and Sensors for Agricultural and Environmental Management
Five of the applications we received intended to use machine learning alongside different types of sensors for use cases relating to agricultural and environmental management.
Applications identified a wide range of different sensors that can be employed to measure different variables within the surrounding environment — including changes to soil quality, ground water levels, moisture levels, temperatures and weather. They also identified different ways in which sensor data might be used alongside Geographic Information Systems (GIS) and a machine learning algorithm to deliver predictive models. These are models where, over time, an algorithm can learn to identify different patterns within data and make new predictions without running pre-programmed queries in order to do so. As more data is consumed the algorithms become more accurate in identifying trends and making predictions.
For the agricultural sector, applicants proposed to develop predictive models to predict the risks of harvest losses due to pest infestation or due to drought, flooding or other meteorological phenomenon. Other applications sought to deliver solutions which could predict the risks of natural disasters to local communities in order to provide time for authorities and communities to act to mitigate risk. Our Call 6 pilot ‘‘Early Warning Forest Fire Detection System’’ is one such example. The pilot proposes to use thermal sensors and machine learning approaches to identify hotspots that are vulnerable to catching fire. This solution will be used as part of an early warning system to prevent and manage wildfires and save forests, communities and associated habitats.
Machine learning and ‘Smart Cities’ in India
Five of the Call 6 applications we received sought to test machine learning related solutions for improving different aspects of city life in India.
Applicants explored use cases related to the transport sector and transport planning. This included proposals to use analytical models to help planners to predict traffic congestion and identify where interventions are most needed. Applicants also sought to use machine learning alongside cameras for crowd management (particularly in congested areas like train stations).
Alongside transport related use cases, applicants also explored how machine learning might be used to support the optimisation of energy distribution and consumption within cities. They also proposed to develop analytical models which could help public officials pinpoint locations within cities where public safety interventions were most needed.
This relatively high number of machine learning applications for city related use cases is to some extent symptomatic of wider trends within India where urbanisation has surged in recent years with up to 45% of the population now living in urban areas. Under initiatives such as the ‘Smart Cities Mission’ the national government has developed plans for piloting and testing technologies (including AI and machine learning) in at least 99 different cities to address infrastructural and service level needs within cities. A wide variety of different public, private and not-for-profit machine learning pilots have been carried out in cities across India, including solutions for monitoring air quality, public safety and electricity theft.
‘Data hungry machines’ and the need for ‘good’ data
Many of our Call 6 machine learning applications identified the important role data would play in their proposed pilot. Applicants told us that machine learning is often a ‘data hungry’ process and that machine learning algorithms often need to be fed large quantities of data, in the right format, in order to learn to spot trends and patterns.
Many applicants proposed plans for accessing historical data sets in order to ‘feed’ solutions with the information needed to learn. They identified key partnerships with ecosystem actors who might be able to provide the data required for solutions.
To ensure consistent, high quality data, some applications also put forward plans for cleaning data into formats that could then be used to teach machines. Here cleaning of data meant anything from formatting disparate data sets so that they conformed to similar structures and taxonomies, to ensuring photographic data was clear (and not blurry or out of focus).
Common problems occurring with drone imagery for Machine Learning — as experienced by our ML for Road Conditions Analysis pilot: LEFT — high blurring, MIDDLE — overexposure, RIGHT — a high quality image that was useable for machine learning (Photo Credit: ML for Road Conditions Analysis)
These types of considerations around data have been highlighted by previous Frontier Technologies Hub machine learning pilots. Our pilot on Machine Learning for Road Conditions Analysis in Zanzibar previously published a blog post on the challenges they experienced accessing and structuring data into the format they needed to train a machine learning algorithm. The blog post shares their experience, which included fixing data sets by hand. It also shares some wider lessons for any pilot looking to conduct machine learning for international development projects — including the lesson for pilot teams to act as early as possible and be more involved in accessing and processing the data needed for machine learning.
Placing limits on machine learning, identifying the risks and value add
A number of Call 6 applications demonstrated an awareness for where they might place limits on the use and role of machine learning as part of a wider solution. This included recognising where they felt machine learning should not be used as a substitute for human expertise and judgement. Applicants also recognised what approaches would (and would not) be acceptable to stakeholders. Most of the machine learning applications we received did not try to fully automate end to end processes. Many sought to provide end users like farmers or environmental authorities with better information to make decisions, through solutions like online dashboards or SMS based advisory services.
Considerations around the role of machine learning within broader solutions have been explored by previous Frontier Technologies pilots. Our previous pilot ‘Artificial Intelligence for TB’ tested a machine learning solution capable of identifying the presence of TB and silicosis in chest X-rays. During the pilot, the implementing team recognised that, at this stage, the tool was unable to differentiate TB from silicosis in X-rays to a high degree of sensitivity and specificity. Consequently the pilot concluded that human medical expertise was still required in the diagnosis process. The AI for TB team also found that by being less ambitious on what their machine learning should accomplish, they were able to establish more buy-in from stakeholders and a quicker path towards adoption and scaling.
Since the completion of the pilot, the AI for TB team have undertaken a detailed exploration on the different ways in which the introduction of ‘disruptive’ technologies like machine learning can lead to distrust from stakeholders. They have identified the risks and ethical implications that machine learning has potential to create. This includes the risk to privacy (where personal information is generated, shared and misused without consent), or wider risks to the supporting ecosystems. For AI for TB, this meant risks of undermining transparency, accountability, accuracy and human skills within health systems. The AI for TB team have written a journal article, using AI for TB as a case study, to explore these different risks and potential mitigations.
A number of Call 6 machine learning applications explored the potential risks of their proposed solutions. Some applications acknowledged that their solution would capture personal data, and identified the need for the pilot to explore the range of risks (such as risks to privacy and risks of surveillance) that would need to be mitigated. Likewise some of the Smart Cities related applications acknowledged an additional risk of their solution entrenching existing biases within cities. They acknowledged that in order to deliver equitable impact, there was a need for more complete and representative data sets that allowed solutions to highlight areas of greatest need.
While recognising the risks and limits of machine learning solutions, applications were, nevertheless, also able to articulate the value add for adopting machine learning for their use cases. In many cases, this included demonstrating the benefits of a machine learning approach over other approaches such as traditional data analysis and modelling. This included demonstrating the considerable efficiencies or additional insights that machine learning can unlock. These are benefits that former Frontier Technologies pilots have helped to demonstrate. In the case of AI for TB, the pilot demonstrated the potential for machine learning to enable more efficient adjudication of compensation claims from former gold miners with occupational lung disease, in particular, through reducing the amount of input required from (already scarce) health professionals.
In this post we shared the key trend from those Call 6 applications who wished to pilot technologies in South Asia. Look out for our fifth and final blogpost on trends from Call 6, where we will be sharing out a key trend relating to East Africa!Exploring a range of different use cases
