Tag Archives: machine learning

Oops! Satellite Imagery Cannot Predict Human Development Indicators

Guest post from Wayan Vota

In many developing country environments, it is difficult or impossible to obtain recent, reliable estimates of human development. Nationally representative household surveys, which are the standard instrument for determining development policy and priorities, are typically too expensive to collect with any regularity.

Recently, however, researchers have shown the potential for remote sensing technologies to provide a possible solution to this data constraint. In particular, recent work indicates that satellite imagery can be processed with deep neural networks to accurately estimate the sub-regional distribution of wealth in sub-Saharan Africa.

Testing Neural Networks to Process Satellite Imagery

In the paper, Can Human Development be Measured with Satellite Imagery?, Andrew Head, Mélanie Manguin, Nhat Tran, and Joshua Blumenstock explore the extent to which the same approach – of using convolutional neural networks to process satellite imagery – can be used to measure a broader set of human development indicators, in a broader range of geographic contexts.

Their analysis produces three main results:

  • They successfully replicate prior work showing that satellite images can accurately infer a wealth-based index of poverty in sub-Saharan Africa.
  • They show that this approach can generalize to predicting poverty in other countries and continents, but that the performance is sensitive to the hyperparameters used to tune the learning algorithm.
  • They find that this approach does not trivially generalize to predicting other measures of development such as educational attainment, access to drinking water, and a variety of health-related indicators.

This paper shows that while satellite imagery and machine learning may provide a powerful paradigm for estimating the wealth of small regions in sub-Saharan Africa, the same approach does not trivially generalize to other geographical contexts or to other measures of human development.

In this assessment, it is important to emphasize what they mean by “trivially,” because in truth the point they are making is somewhat circumspect. Specifically, what they have shown is that the exact framework—of retraining a deep neural network on night-lights data, and then using those features to predict the wealth of small regions in sub-Saharan Africa—cannot be directly applied to predicting arbitrary indicators in any country with uniformly good results.

This is an important point to make because absent empirical evidence to the contrary, it is likely that policymakers eager to gain quick access to micro-regional measurements of development might be tempted to do exactly what they have done in this paper, without paying careful attention to the thorny issues of generalizability that they have uncovered in this analysis.

It is not the researchers’ intent to impugn the potential for related approaches to provide important new methods for measuring development, but rather to say that such efforts should proceed with caution, and with careful validation.

Why Satellite Imagery Might Fail to Predict Development

The results showed that while some indicators like wealth and education can be predicted reasonably well in many countries, other development indicators are much more brittle, exhibiting high variance between and within countries, and others perform poorly everywhere.

Thus it is useful to distinguish between two possible reasons why the current approach may have failed to generalize to these measures of development.

  • It may be that this exercise is fundamentally not possible, and that no amount of additional work would yield qualitatively different results.
  • It is quite possible that their investigation to date has been not been sufficiently thorough, and that more concerted efforts could significantly improve the performance of these models

Insufficient “signal” in the satellite imagery.

The researchers’ overarching goal is to use information in satellite images to measure different aspects of human development. The premise of such an approach is that the original satellite imagery must contain useful information about the development indicator of interest. Absent of such a signal, no matter how sophisticated our computational model, the model is destined to fail.

The fact that wealth specifically can be measured from satellite imagery is quite intuitive. For instance, there are visual features one might expect correlate with wealth—large buildings, metals roofs, nicely paved roads, and so forth.

It may be the case that other measures of human development cannot be seen from above. For instance, it may be a fundamentally difficult task to infer the prevalence of malnutrition from satellite imagery, if the regions with high and low rates of malnutrition appear similar, even though they hypothesize that these indices should correlate with wealth index.

They were, however, surprised by the relative under-performance of models designed to predict access to drinking water, as they expected the satellite-based features to capture proximity to bodies of water, which in turn might affect access to drinking water.

(Over-) reliance on night-lights may not generalize.

Their reliance on night lights might help explain why some indicators were predicted less successfully in some countries than others. An example in their study includes Nepal, where the accuracy in predicting access to electricity was much lower (R2 = 0.24) than in the other countries (R2 = 0.69, 0.44, and 0.54 in Rwanda, Nigeria, and Haiti, respectively).

This may be partly due to the fact that Nepal has a very low population density (half as dense as Haiti and Rwanda) and very high levels of electrification (twice as high as Haiti, Rwanda, and Nigeria).

If the links between electrification, night-lights, and daytime imagery are broken in Nepal, they would expect their modeling approach to fail. More generally, they expect that when a development indicator does not clearly relate to the presence of nighttime lights, it may be unreasonable to expect good performance from the transfer learning process as a whole.

Deep learning vs. supervised feature engineering.

In this paper, the researchers focused explicitly on using the deep/transfer learning approach to extracting information from satellite images. While powerful, it is also possible that other approaches to feature engineering might be more successful than the brute force approach of the convolutional neural network.

For instance, Gros and Tiecke have recently shown how hand-labeled features from satellites, and specifically information about the types of buildings that are present in each image, can be quite effective in predicting population density. Labeling images in this manner is resource intensive, and they did not have the opportunity to test such approaches.

However, they believe that careful encoding of the relevant information from satellite imagery would likely bolster the performance of specific prediction tasks.

Neural Networks Can Still Process Satellite Imagery

Broadly, the researchers remain optimistic that future work using novel sources of data and new computational algorithms can engender significant advances in the measurement of human development.

However, it is imperative that such work proceeds carefully, with appropriate benchmarking and external calibration. Promising new tools for measurement have the potential to be implemented widely, possibly by individuals who do not have extensive expertise in the underlying algorithms.

Applied blindly, these algorithms have the potential to skew subsequent policy in unpredictable and undesirable ways. They view the results of this study as a cautionary example of how a promising algorithm should not be expected to work “off the shelf” in a context that is significantly different from the one in which it was originally developed.

The post Oops! Satellite Imagery Cannot Predict Human Development Indicators appeared first on ICTworks.

3 Lessons Learned using Machine Learning to Measure Media Quality

by Samhir Vasdev, Technical Adviser for Digital Development at IREX’s Center for Applied Learning and Impact. The post 3 Lessons Learned using Machine Learning to Measure Media Quality appeared first on ICTworks.

Moving from hype to practice is an important but challenging step for ICT4D practitioners. As the technical adviser for digital development at IREX, a global development and education organization, I’ve been watching with cautious optimism as international development stakeholders begin to explore how artificial intelligence tools like machine learning can help them address problems and introduce efficiencies to amplify their impact.

So while USAID was developing their guide to making machine learning work for international development and TechChange rolled out their new course on Artificial Intelligence for International Development, we spent a few months this summer exploring whether we could put machine learning to work to measure media quality.

Of course, we didn’t turn to machine learning just for the sake of contributing to the “breathless commentary of ML proponents” (as USAID aptly puts it).

As we shared in a session with our artificial intelligence partner Lore at MERLTech DC 2018, some of our programs face a very real set of problems that could be alleviated through smarter use of digital tools.

Our Machine Learning Experiment

In our USAID-funded Media Strengthening Program in Mozambique, for example, a small team of human evaluators manually score thousands of news articles based on 18 measures of media quality.

This process is time consuming (some evaluators spend up to four hours a day reading and evaluating articles), inefficient (when staff turns over, we need to reinvest resources to train up new hires), and inconsistent (even well-trained evaluators might score articles differently).

To test whether we can make the process of measuring media quality less resource-intensive, we spent a few months training software to automatically detect one of these 18 measures of media quality: whether journalists keep their own opinions out of their news articles. The results of this experiment are very compelling:

  • The software had 95% accuracy in recognizing sentences containing opinions within the dataset of 1,200 articles.
  • The software’s ability to “learn” was evident. Anecdotally, the evaluation team noticed a marked improvement in the accuracy of the software’s suggestions after showing it only twenty sentences that had opinions. The accuracy, precision, and recall results highlighted above were achieved after only sixteen rounds of training the software.
  • Accuracy and precision increased the more that the model was trained. There is a clear relationship between the number of times the evaluators trained the software and the accuracy and precision of the results. The recall results did not improve over time as consistently.

These results, although promising, simplify some numbers and calculations. Check out our full report for details.

What does this all mean? Let’s start with the good news. The results suggest that some parts of media quality—specifically, whether an article is impartial or whether it echoes its author’s opinions—can be automatically measured by machine learning.

The software also introduces the possibility of unprecedented scale, scanning thousands of articles in seconds for this specific indicator. These implications introduce ways for media support programs to spend their limited resources more efficiently.

3 Lessons Learned from using Machine Learning

Of course, the machine learning experience was not without problems. With any cutting-edge technology, there will be lessons we can learn and share to improve everyone’s experience. Here are our three lessons learned working with machine learning:

1. Forget about being tech-literate; we need to be more problem-literate.

Defining a coherent, specific, actionable problem statement was one of the important steps of this experiment. This wasn’t easy. Hard trade-offs had to be made (Which of 18 indicators should we focus on?), and we had to focus on things we could measure in order to demonstrate efficiency games using this new approach (How much time do evaluators currently spend scoring articles?).

When planning your own machine learning project, devote plenty of time at the outset—together with your technology partner—to define the specific problem you’ll try to address. These conversations result in a deeper shared understanding of both the sector and the technology that will make the experiment more successful.

2. Take the time to communicate results effectively.

Since completing the experiment, people have asked me to explain how “accurate” the software is. But in practice, machine learning software uses different methods to define “accuracy”, which in turn can vary according to the specific model (the software we used deploys several models).

What starts off as a simple question (How accurate is our software?) can easily turn into a discussion of related concepts like precision, recall, false positives, and false negatives. We found that producing clean visuals (like this or this) became the most effective way to explain our results.

3. Start small and manage expectations.

Stakeholders with even a passing awareness of machine learning will be aware of its hype. Even now, some colleagues ask me how we “automated the entire media quality assessment process”—even though we only used machine learning to identify one of 18 indicators of media quality. To help mitigate inflated expectations, we invested a small amount into this “minimum viable product” (MVP) to prove the fundamental concept before expanding on it later.

Approaching your first machine learning project this way might help to keep expectations in line with reality, minimize risks associated with experimentation, and provide air cover for you to adjust your scope as you discover limitations or adjacent opportunities during the process.

How I Learned to Stop Worrying and Love Big Data

by Zach Tilton, a Peacebuilding Evaluation Consultant and a Doctoral Research Associate at the Interdisciplinary PhD in Evaluation program at Western Michigan University. 
In 2013 Dan Airley quipped “Big data is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it….” In 2015 the metaphor was imported to the international development sector by Ben Ramalingam, in 2016 it became a MERL Tech DC lightning talk, and has been ringing in our ears ever since. So, what about 2018? Well, unlike US national trends in teenage sex, there are some signals that big or at least ‘bigger’ data is continuing to make its way not only into the realm of digital development, but also evaluation. I recently attended the 2018 MERL Tech DC pre-conference workshop Big Data and Evaluation where participants were introduced to real ways practitioners are putting this trope to bed (sorry, not sorry). In this blog post I share some key conversations from the workshop framed against the ethics of using this new technology, but to do that let me first provide some background.
I entered the workshop on my heels. Given the recent spate of security breaches and revelations about micro-targeting, ‘Big Data’ has been somewhat of a boogie-man for myself and others. I have taken some pains to limit my digital data-footprint, have written passionately about big data and surveillance capitalism, and have long been skeptical of big data applications for serving marginalized populations in digital development and peacebuilding. As I found my seat before the workshop started I thought, “Is it appropriate or ethical to use big data for development evaluation?” My mind caught hold of a 2008 Evaluation Café debate between evaluation giants Michael Scriven and Tom Cook on causal inference in evaluation and the ethics of Randomized Control Trials. After hearing Scriven’s concerns about the ethics of withholding interventions from control groups, Cook asks, “But what about the ethics of not doing randomized experiments?” He continues, “What about the ethics of having causal information that is in fact based on weaker evidence and is wrong? When this happens, you carry on for years and years with practices that don’t work whose warrant lies in studies that are logically weaker than experiments provide.”
While I sided with Scriven for most of that debate, this question haunted me. It reminded me of an explanation of structural violence by peace researcher Johan Galtung who writes, “If a person died from tuberculosis in the eighteenth century it would be hard to conceive of this as violence since it might have been quite unavoidable, but if he dies from it today, despite all the medical resources in the world, then violence is present according to our definition.” Galtung’s intellectual work on violence deals with the difference between potential and the actual realizations and what increases that difference. While there are real issues with data responsibility, algorithmic biases, and automated discrimination that need to be addressed, if there are actually existing technologies and resources not being used to address social and material inequities in the world today, is this unethical, even violent? “What about the ethics of not using big data?” I asked myself back. The following are highlights of the actually existing resources for using big data in the evaluation of social amelioration.

Actually Existing Data

During the workshop, Kerry Bruce from Social Impact shared with participants her personal mantra, “We need to do a better job of secondary data analysis before we collect any more primary data.” She challenged us to consider how to make use of the secondary data available to our organizations. She gave examples of potential big data sources such as satellite images, remote sensors, GPS location data, social media, internet searches, call-in radio programs, biometrics, administrative data and integrated data platforms that merge many secondary data files such as public records and social service agency and client files. The key here is there are a ton of actually existing data, many of which are collected passively, digitally, and longitudinally. Despite noting real limitations to accessing existing secondary data, including donor reluctance to fund such work, limited training in appropriate methodologies in research teams, and differences in data availability between contexts, to underscore the potential of using secondary data, she shared a case study where she lead a team to use large amounts of secondary indirect data to identify ecosystems of modern day slavery at a significantly reduced cost than collecting the data first-hand. The outputs of this work will help pinpoint interventions and guide further research into the factors that may lead to predicting and prescribing what works well for stopping people from becoming victims of slavery.

Actually Existing Tech (and math)

Peter York from BCT Partners provided a primer on big data and data science including the reality-check that most of the work is the unsexy “ETL,” or the extraction, transformation, and loading of data. He contextualized the potential of the so-called big data revolution by reminding participants that the V’s of big data, Velocity, Volume, and Variety, are made possible by the technological and social infrastructure of increasingly networked populations and how these digital connections enable the monitoring, capturing, and tracking of ever increasing aspects of our lives in an unprecedented way. He shared, “A lot of what we’ve done in research were hacks because we couldn’t reach entire populations.” With advances in the tech stacks and infrastructure that connect people and their internet-connected devices with each other and the cloud, the utility of inferential statistics and experimental design lessens when entire populations of users are producing observational behavior data. When this occurs, evaluators can apply machine learning to discover the naturally occurring experiments in big data sets, what Peter terms ‘Data-driven Quasi-Experimental Design.’ This is exactly what Peter does when he builds causal models to predict and prescribe better programs for child welfare and juvenile justice to automate outcome evaluation, taking cues from precision medicine.
One example of a naturally occurring experiment was the 1854 Broad Street cholera outbreak in which physician John Snow used a dot map to identify a pattern that revealed the source of the outbreak, the Broad Street water pump. By finding patterns in the data John Snow was able to lay the groundwork for rejecting the false Miasma Theory and replace it with a proto-typical Germ Theory. And although he was already skeptical of miasma theory, by using the data to inform his theory-building he was also practicing a form of proto-typical Grounded Theory. Grounded theory is simply building theory inductively, after data collection and analysis, not before, resulting in theory that is grounded in data. Peter explained, “Machine learning is Grounded Theory on steroids. Once we’ve built the theory, found the pattern by machine learning, we can go back and let the machine learning test the theory.” In effect, machine learning is like having a million John Snows to pour over data to find the naturally occurring experiments or patterns in the maps of reality that are big data.
A key aspect of the value of applying machine learning in big data is that patterns more readily present themselves in datasets that are ‘wide’ as opposed to ‘tall.’ Peter continued, “If you are used to datasets you are thinking in rows. However, traditional statistical models break down with more features, or more columns.” So, Peter and evaluators like him that are applying data science to their evaluative practice are evolving from traditional Frequentist to Bayesian statistical approaches. While there is more to the distinction here, the latter uses prior knowledge, or degrees of belief, to determine the probability of success, where the former does not. This distinction is significant for evaluators who are wanting to move beyond predictive correlation to prescriptive evaluation. Peter expounded, Prescriptive analytics is figuring out what will best work for each case or situation.” For example, with prediction, we can make statements that a foster child with certain attributes is 70% not likely to find a home. Using the same data points with prescriptive analytics we can find 30 children that are similar to that foster child and find out what they did to find a permanent home. In a way, only using predictive analytics can cause us to surrender while including prescriptive analytics can cause us to endeavor.

Existing Capacity

The last category of existing resources for applying big data for evaluation was mostly captured by the comments of independent evaluation consultant, Michael Bamberger. He spoke of the latent capacity that existed in evaluation professionals and teams, but that we’re not taking full advantage of big data: “Big data is being used by development agencies, but less by evaluators in these agencies. Evaluators don’t use big data, so there is a big gap.”

He outlined two scenarios for the future of evaluation in this new wave of data analytics: a state of divergence where evaluators are replaced by big data analysts and a state of convergence where evaluators develop a literacy with the principles of big data for their evaluative practice. One problematic consideration with this hypothetical is that many data scientists are not interested in causation, as Peter York noted. To move toward the future of convergence, he shared how big data can enhance the evaluation cycle from appraisal and planning through monitoring, reporting and evaluating sustainability. Michael went on to share a series of caveats emptor that include issues with extractive versus inclusive uses of big data, the fallacy of large numbers, data quality control, and different perspectives on theory, all of which could warrant their own blog posts for development evaluation.

While I deepened my basic understandings of data analytics including the tools and techniques, benefits and challenges, and guidelines for big data and evaluation, my biggest take away is reconsidering big data for social good by considering the ethical dilemma of not using existing data, tech, and capacity to improve development programs, possibly even prescribing specific interventions by identifying their probable efficacy through predictive models before they are deployed.

(Slides from the Big Data and Evaluation workshop are available here).

Do you use or have strong feelings about big data for evaluation? Please continue the conversation below.