CIT computer scientist Milan Cvitkovic conducted 46 in-depth interviews with "scientists, engineers, and CEOs" and collated their machine learning research needs into an aptly named paper entitled "Some Requests for Machine Learning Research from the East African Tech Scene," which presents an illuminating look into the gaps in the current practice of machine learning, itself an example of how rich-world priorities shape our ability to understand, compute and predict the world.
Some of the gaps are predictable enough (regional languages are underrepresented in speech-to-text models) and others are somewhat surprising (speech-to-text models are really bad at recognizing when speakers "code switch" between languages mid-sentence, which is a common practice in the region) and some are really thorny (due to regional "low trust" economies, "interviewees who use machine learning with surveys or customer interaction data reported spending significant effort fighting fraud or dishonesty").
Reinforcement Learning – No interviewee reported using any reinforcement learning methods. However, interest was expressed in it, particularly regard ing machine teaching and using RL in simulations, e.g. using RL in epidemiological simulations to find worst case scenarios in outbreak planning.
Machine Teaching – There is a shortage of good educational resources and teachers in East Africa. Several initiatives exist that use mobile phones as an education platform. Practitioners were interested in using ideas from machine teaching in their work to personalize content delivered. However, the author did not encounter anyone who had employed any results from the machine teaching literature at this point.
Uncertainty Quantification – An important factor that keeps the wealth of rich regions from moving into poorer regions like East Africa, despite the fact that it should earn greater returns there, is risk . Not all risk can be machine–learned away by any means. But (accurate) predictive models are risk-reduction tools.
Machine learning models are most useful for risk–reduction when they can (accurately) quantify their uncertainty. This is particularly true when data are scarce, as they usually are in East Africa. UQ is not a new problem by any means, but it is listed here to reiterate its importance to the organizations interviewed. Importantly, when used in East Africa, UQ is typically much more concerned with conservatively quanti
fying overall downside risk (with respect to some quantity of interest) than characterizing overall model uncertainty around point predictions.
Some Requests for Machine Learning Research from the East African Tech Scene [Milan Cvitkovic/Arxiv]
(via Four Short Links)
(Image: Cryteria, CC-BY)