Reading:
Excavating AI: The Politics of Images in Machine Learning Training Sets by Kate Crawford and Trevor Paglen
Quotes that stood out:
Prompt:
Reflect on the relationship between labels and images in a machine learning image classification dataset. Who has the power to label images and how do those labels and machine learning models trained on them impact society?
Response:
Algorithms trained to classify images don’t always get the labels right. Society is constantly evolving—new perceptions form daily, new trends emerge weekly, and change happens over time. While labels can attempt to keep pace, human behavior in a rapidly shifting world will always stay a step ahead. Machine learning systems can adapt, but the nuances of human connection—our "inside jokes" and shared experiences that machines cannot easily interpret—will often go unrecognized.
I believe we, as humans, hold the power to label images. These labels are a reflection of our perspectives and the society we live in, shaped by our shifting views, economy, and trends. As Stanford Professor Fei-Fei Li stated, the idea behind ImageNet was to “map out the entire world of objects.” But the act of labeling goes beyond mere categorization—it influences how we see and understand the world around us.
In 1839, mathematician François Arago claimed that through photographs, “objects preserve mathematically their forms.” When placed in the context of 19th-century imperialism and social Darwinism, photography became a tool that lent a “scientific” veneer to practices like phrenology, physiognomy, and eugenics. This historical misuse of imagery to reinforce harmful ideologies underscores the power and impact of labeling in visual datasets.
It’s fascinating to consider that, at their core, images are just pixels—binary data that preserves a moment in real life mathematically. This idea resonates with the concept that images "preserve" forms, whether as pixels or as part of human memory. Yet, how we label these images is also a reflection of our deeper understanding of concepts like age, gender, and race.
I think it’s also interesting how images are labeled:
There’s so much complexity in how we, as humans, interpret these characteristics. Age brings with it different stages of life—grandparents, parents, teens, babies— not just a number. Gender is not just a binary but an identity, and race is an intricate mixture of history and experience. These are delicate and deeply personal topics, shaped by human understanding and self-representation. A machine, however, doesn’t grasp these layers. It reduces the world to 1s and 0s, to black or white, and in doing so, risks oversimplifying the very essence of human experience.
Racial recognition, in particular, underscores the importance of a programmer’s understanding of diversity and inclusivity. As highlighted in the article, IBM’s “Diversity in Faces” dataset was developed in response to critics who revealed that the company’s facial-recognition software frequently failed to recognize people with darker skin. This highlights a crucial point: an algorithm cannot exhibit inclusivity if it's not trained on a diverse dataset. It’s akin to teaching a preschooler to read—if you never introduce them to phonetics, how can they ever learn? Similarly, if a programmer neglects to include diverse representations in the training process, the fault lies with them, not the algorithm, for failing to be inclusive.