Reading:

[It] assumes that concepts… exist in the first place, and that those concepts are fixed, universal, and have some sort of transcendental grounding and internal consistency. It assumes a fixed and universal correspondences between images and concepts, appearances and essences.

Human language (specifically in this context, English) is complicated and varied, with historical contexts and regional variations and connotations that constantly shift and evolve. Limiting the data set to just nouns doesn’t change any of that. Additionally, not all nouns can be represented purely visually—characteristics relating to people, especially, are too complicated to be accurately represented with a photograph.

In the case ImageNet, the labeling was outsourced to masses on Amazon Mechanical Turk, resulting in a dataset that potentially holds all the many varied biases and judgements of the crowd. However, the alternative is that a select group of like-minded people label the images, which would in theory be even more “opinionated.” But these labels are also picked from online dictionaries, with outdated language and missing important historical context.

The ImageNet classification hierarchy recalls the old Library of Congress classification of LGBTQ-themed books under the category “Abnormal Sexual Relations, Including Sexual Crimes,” which the American Library Association's Task Force on Gay Liberation finally convinced the Library of Congress to change in 1972 after a sustained campaign

For all its abilities in pattern recognition, machine learning has not—and maybe can not— master the nuance of human language.


Coding Exercise

Screen Recording 2022-09-23 at 1.30.33 AM.mov

Edit code OR fullscreen view

Process:

I just wanted to start with something relatively simple since I forgot a lot of p5.js over the summer, and for this idea all that was really necessary was to make a circle move over the screen.

I tested the directions with keyPressed() first just to make sure everything was working, and then went to Teachable Machine with pretty basic arm movements for 4 directions.

Screen Shot 2022-09-23 at 1.01.47 AM.png

After that was basically just doing what we did in class and replacing keyPressed() with the labels from the model.