­What Gets Lost When We Design for the Average?

Andrei Mihai

Three minutes is not much time to explain a research project. Yet that was the challenge facing the young researchers taking the stage during the Next Gen session at the 13th Heidelberg Laureate Forum in September. One after another, they moved through problems in mathematics and computer science: quantum systems, knot theory, weather prediction, brain imaging, genomics, and artificial intelligence.

The subjects changed quickly, jumping from theme to theme. But across a few of the presentations, a common thread emerged.

Science and technology often become useful at scale by simplifying and classifying things. You can standardize a language or treat a population as a dataset. You can analyze average health changes. Those simplifications are often necessary, but what do they leave out?

For some of the young researchers the answer was: quite a lot.

When AI Speaks Your Language, But Not Your Dialect

Priscilla Chinwendu Ezulu began with an apparently straightforward question. What happens when AI learns a version of a language, but many people also use other dialects?

More than 30 million people speak Igbo, Ezulu told the audience, but there is no single uniform way of speaking it. The language contains more than 20 dialects, with differences in pronunciation, tone and vocabulary.

AI systems, however, tend to work with a streamlined version they just call “Igbo.” For millions of people, that just won’t cut it.

Different Igbo dialects can express the same idea in noticeably different ways. To a human speaker, they belong to the same linguistic landscape, but to a machine trained on a narrower version of that language, they may look like entirely different inputs.

Her ongoing project, “Dialect-Aware Neural Models for Low-Resource African Languages: A Case Study of Igbo Language,” aims to make AI systems more sensitive to that variation. At the HLF, she described plans to gather data from at least 7,000 speakers, amounting to roughly 350 hours of speech, and to work with Igbo teachers to validate the recordings.

The goal is to enable AI systems to recognize which variety of Igbo somebody is actually speaking and then switch to that version. Speech and language are some of the first widespread AI applications, but research on low-resource languages has repeatedly identified the lack of diverse training data as a major obstacle for speech-recognition systems. Ezulu’s work aims to tackle that problem. AI, she argued, should adapt to linguistic diversity rather than assume uniformity.

A School Is More Than a Point on the Map

For Natalia Garcia San Miguel, the problem of false uniformity appears in a completely different setting: menstrual-product access in Peru.

A common way to design social interventions is geographical. Identify an area with a particular problem, group together the institutions within that area, and devise a solution. But two schools that look close together on a map may be dramatically different in practice.

Garcia San Miguel’s work considers not just geography but infrastructure, water access, school characteristics and social conditions. Some of those variables can be measured straightforwardly, like ratios of students to teachers or toilets, for instance. Others are categorical or social, like urban versus rural, or the prevalence of stigma around menstruation.

Taken together, they produce a much more detailed picture of the environment in which an intervention actually has to work. She then uses those variables to identify groups of schools with genuinely similar needs rather than simply assuming nearby schools are alike.

Behind this mathematical problem is a very tangible initiative. Garcia San Miguel is involved with MenstruAcción, a youth-led initiative working on menstrual dignity and education in Peru, including projects in Lima, Cajamarca, and Amazonas.

The organization treats menstrual poverty as more than the absence of pads or other products. Infrastructure, information, privacy, health, economic conditions, and social attitudes all matter. The mathematics García San Miguel presented at the Forum in Heidelberg tries to capture some of that complexity.

Your Smartwatch Has Never Met the Average Person

Lidia Alecci brought the same problem all the way down to the individual body.

Smartphones and smartwatches now collect extraordinary amounts of information about their users. Machine-learning models can use sensor data to estimate activity, sleep, stress, and other aspects of our physical or psychological state. Typically, Alecci explained, those systems learn from data collected from many people. The result is a population model: patterns that work reasonably well across a group.

But people vary tremendously. They differ physiologically, they experience emotions differently, their bodies and routines change with age, health, and circumstances. Even the same person is not necessarily consistent from one month to the next. So then, how useful is a model trained on an average population when it encounters one particular human being?

Alecci’s PhD research asks when these differences matter enough that a population model should give way to a personalized one.

In one study, she and her colleagues analyzed 6,557 hours of sensor data collected from 16 participants wearing wrist devices for a month. Personalized models were significantly better than population models at predicting subjective sleep quality, while personalization offered much less advantage for the more objective task of distinguishing sleep from wakefulness.

Personalization can offer better data, but it creates problems of its own. Learning an individual requires information about that individual, and collecting enough labelled personal data can be expensive, inconvenient, or raise privacy questions. So, Alecci’s research is not simply arguing that personalized models are always better.

The more interesting question is when individuality matters enough to justify personalization – and what it costs to capture it.

Many Different Types of Data

The Next Gen session went considerably further than these projects. Another presentation showed how a machine-learning system that appeared better at detecting nerve-fiber bundles in the brain could actually produce less accurate outlines – a reminder that the metric used to judge a model can conceal important kinds of errors. Another asked how neural networks could become interpretable enough to help physicists understand the patterns they discover, rather than simply operating as highly successful black boxes.

But the issue of data came up over and over, not just in the Next Gen session, but also in the subsequent Poster one.

Xie Zhang, from the University of Hong Kong, presented research on using thermal sensors to turn buildings into what he calls cognizant spaces.

Rather than relying on conventional cameras, his systems use arrays of thermal sensors to capture the heat emitted by human bodies. Combined with machine learning, these measurements can be used to reconstruct three-dimensional body geometry, recognize gestures, and monitor movement.

The approach has already been explored in applications ranging from fall detection to sleep-posture monitoring. Zhang’s poster also described deployments in offices, gyms, and university facilities, where the technology could operate continuously without capturing conventional, recognizable images of the people being monitored. Personalization also comes in the form of body heat, apparently.

But body heat is not the only information a building can collect.

In another poster, Mainak Chakraborty and Subrat Kar demonstrated how structural vibrations could reveal surprisingly detailed information about the people walking across a floor.

Every footstep produces a small mechanical wave that travels through the building. A geophone placed beneath the floor can record those vibrations, while machine-learning models analyze the signals to infer who is walking, how they are moving and potentially even whether they are distracted.

Their research explores applications in mobility assessment, fall-risk screening, and remote patient monitoring.

Yet these technologies also introduce a privacy concern. You can remove conventional cameras and still not necessarily eliminate surveillance. A floor that recognizes individual walking patterns, or a thermal sensor that continuously tracks human movement, can still reveal information about people without ever recording their faces.

Diversity in Research

There were many other talks and posters on making numerical weather prediction more stable, performing genomic analysis closer to where enormous datasets are stored, modelling networks of fibers in materials such as paper, and mathematical journeys through graph theory, additive combinatorics, algebraic geometry, and knot theory.

In that sense, there was no single message to the Next Gen and Poster sessions – nor should there have been. The purpose was to offer a glimpse of the questions occupying a new generation of mathematicians and computer scientists.

But some of the most compelling talks themselves shared a particular resistance to convenient categories. There is no single way to speak Igbo, schools in the same area can face the very different constraints, and a model trained on thousands of people may still misunderstand the individual wearing the watch.

Scale almost always demands simplification.

The post ­What Gets Lost When We Design for the Average? originally appeared on the HLFF SciLogs blog.