My Work

<aside> 💡 1. Read Excavating AI: The Politics of Images in Machine Learning Training Sets by Kate Crawford and Trevor Paglen. Consider the following excerpt from the conclusion:

The artist René Magritte completed a painting of a pipe and coupled it with the words “Ceci n’est pas une pipe.” Magritte called the painting La trahison des images, “The Treachery of Images.”

Magritte’s assumption was almost diametrically opposed: that images in and of themselves have, at best, a very unstable relationship to the things seem to represent, one that can be sculpted by whoever has the power to say what a particular image means. For Magritte, the meaning of images is relational, open to contestation. At first blush, Magritte’s painting might seem like a simple semiotic stunt, but the underlying dynamic Magritte underlines in the painting points to a much broader politics of representation and self-representation.

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?

</aside>

Who has the power?

There’s a difference between who currently has the power and and who should have the power to label images for machine learning models. In the past and to some degree now, anyone who so chooses to and who has the technology to do so can gather images, give them labels, and use them to train ML models.

Despite the fact that datasets are now being taken down, As the text discussed, just the act of removing something from public eye, does not mean they are not still being used in established societal procedures (such as airport security). Moreover, this also does not mean large companies are not still creating and using problematic datasets behind the scenes.

On the other hand, regarding who should have the power to label mages and train ML models, there are a lot of factors to consider including: who these models will serve, what purpose they have, and the consent of subjects / owners of the images.

I think it’s also important to remember that even if all these factors are considered, the labels attributed to images, items and people can still be inaccurate, and should be rigorously subject to review as society, impacted people and purpose of datasets change.

How does it impact society?

As discussed in the reading, all acts of labelling images, creating ML models and datasets, is political. The assumptions made at every step say something about the decisions maker’s stance on certain issues. Here are some practices that influence a lot of ML vision functions currently in practice:

• For example, in the UTKFace dataset and IBM’s ‘improved’ ‘Diversity in Faces’ dataset, there were gender was a binary classification of: ‘male’ and ‘female.’
• The practice of mapping a person’s essential character to features of their faces and bodies and physical attributes.
• categorizing humans into (often harmful and untrue) stereotyped groups.

As things currently stand, once an image or dataset is posted online, it’s reasonable to assume that others might use this data in ways that are impossible to control. I am not saying whether or not this is correct, but that we should assume that this will happen.

In fact, as the article we read this week states “The training sets of labeled images that are ubiquitous in contemporary computer vision and AI are built on a foundation of unsubstantiated and unstable epistemological and metaphysical assumptions about the nature of images, labels, categorization, and representation.”

And when this does happen, harmful stereotypes such as race and gender are perpetuates. It can also cause unfair action to be inspired. A good example of this is using predictive algorithms to detect criminal activity.

<aside> 💡 Train your own image classifer using transfer learning and ml5.js and apply the model to an interactive p5.js sketch. You can train the model with Teachable Machine or with your own ml5.js code. Feel free to try sound instead of or in addition to images. You may also choose to experiment with a "regression" rather than classification.

</aside>