The Platform

Photo illustration by John Lyman

In May, a research paper exposed a hidden bias in how computer-generated images categorized persons with even small amount of mixed genes to lower socioeconomic status. But AI researchers and engineers are not the only ones challenged when building categorization models. The question of how to categorize the world has intrigued humans for centuries.

Categorization is putting items, events, and people who share common traits into groups. We all perform this task to understand and judge the world around us. Over the centuries, philosophers and scientists grappled with the question of finding the ideal way to categorize the world. They invented different frameworks to guide our thinking, and these frameworks sometimes complement one other, and at other times are sharply contradicted in their assumptions and limitations.

Aristotle invented the classic view of categorization, now known as the traditional view. He held that each category should be fully defined, and items should be either inside or outside. The stoic also had four entities: substance, quality, disposition, and relative disposition.

Centuries later, the Austrian philosopher Ludwig Wittgenstein questioned the previous categorizations. He proposed that items inside the category do not need to share one unified property and could instead be connected through overlapping features that he called family resemblance.

Wittgenstein also noticed that the category boundaries could be extended and not fixed. In another substantial departure from the classic view, Wittgenstein argued that some elements are more central than others when we define any category.

In what could be considered a significant shift in our understanding of categorization, the well-known mathematician Lutfi Zadeh introduced the notion of fuzzy sets. He argued that members of sets might have a degree of membership as opposed to the classic view that a member either belongs or does not to a particular set. This shift had a significant impact on real-life applications.

Folk taxonomy classifies objects using common knowledge. Examples of folk taxonomies are found for plants and animals. This categorization usually changes over time and conflicts with general scientific categorization.

Ordinary individuals also categorize the world by organizing items to serve a particular purpose. This categorization is not intended to be shared or influence the public; it follows its own logic and rationale. An example of this type of categorization is organizing personal papers and emails.

Scientists introduced many theories that explain how humans learn concepts such as the rule-based theory that postulates categorization could be realized through setting rules.

If we have two groups, one for square objects and the other for round ones, the rule will be “first category is for square objects.” Exemplar theory assumes that humans learn new objects by recalling an exemplar already stored in the brain. In contrast, the prototype theory argues that humans utilize the prototype to learn recent categorization decisions. The explanation-based theory argues that humans learn concepts by experiencing new examples and forming a new outline.

Some researchers recently argued that categorization happens in a Bayesian fashion in that humans first generate probabilities for a particular concept definition based on previous examples they encountered.

Categorization is a salient issue; it’s the lens through which we see the world. Our prolonged history as humans in categorizing the world reveals that we have always wanted more than one framework that describes how the world could be put into groups to understand how it works. We should, however, utilize these frameworks we build and know their limitations.

Mohamed Suliman is a senior researcher at Northeastern University and also holds a degree in Engineering form the University of Khartoum.