Two problems:
- Lack of expert labor
- Lack of actionable data
Data Science Africa conference - tries to describe unique problems in Africa. Potential uses and datasets Agriculture
- disease monitoring - takes about 6 months to turn surveillance into actionable map. ML can make this intelligence arrive in near real-time.
- Farmers take a picture of their garden, upload the picture, unlabeled.
- They tried asking farmers for labels, but it was mostly to calibrate how well they knew the diseases.
- Incentivizing farmers to do this is tricky - they need data airtime
- Attempts to incentivize: rockstar, ranking them ("this guy is a rockstar!”), social workers
- Diagnosis of leaf images (disease, severity)
- Example tasks:
- Automated whitefly count. There can be hundreds of them on a leaf, you need to take several samples - task is infuriating.
- Necrosis measurement of roots
Human Health
- Automated lab diagnostics
- Many require phone/microscope attachments - 3d printing these is cheap
- Example tasks:
- Automate diagnosis of malaria from blood slides/microscope. Typically you count them by hand, but there are more microscopes than experts. Solution is to put it on a mobile phone (through the eyepiece of the microscope). But this is currently done with thick blood slides, accuracy is in the high 90s; thin blood slides would be an improvement.
- Tuberculosis bacili in sputum
- Worms in fecal matter
Radio mining: "the real social media of Uganda"
- People calling in and reporting things that happened or airing grievances
- Capture data with an antenna and a microphone
Telecoms data:
- Aggregate statistics of movement based on Call Data Records (CDR)
- Hard to get this data, private, so they use aggregated & anonymized data.
- Example usage:
- Understanding a typhoid outbreak - where the epicenter was, how it moved.
Poverty and the environment with satellite imagery
- Tasks:
- Car counting to estimate populations for planning purposes. This is more accurate than counting houses.
Thoughts on deployment/challenges
- Low expectations
- Don’t come in expecting to change the world
- High expectations
- Dunning-Kruger effect: people can really overestimate ML’s power once they get a taste of it
- Pilotitis: they do endless pilots
- Data issues: access, representativeness, privacy
- Theory-practice tradeoff: DL is often too expensive in practice
- Quality control
- Bureaucracy
Final note: collaborator/stakeholder involvement is key: efforts without these never work
Scale of these efforts: there are maybe 5 students at the university in Kumpala, Uganda; maybe 8 in another university (Jakarta?)