Excavating AI: The Politics of Images in Machine Learning Training Sets Reading

• I was quite baffled at the numerous examples of outrageous labels given given to images without warrant or need. I had never realized that biases and prejudice were so prevalent in training datasets.

  • My curiosity got the best of me and I went looking myself for some outrageous labels in the ImageNet sample images. While the github seems to only host innocuous labels of animals and fruits I did find something that did offend me, as an italian. Below is an image, part of the imagenet-sample-images, labeled “carbonara.” Now, this is not in any way close to as outrageous as the labels and corresponding images described in the article, however the fact that this is the reference image for carbonara is blasphemous against italian culture.

    

TensorFlow.js + COCO and ml5 + MobileNet object detection from video

• I played around a little bit with this sketch and found that it was pretty much only good at detecting people and everything else it detected was wrong or just very unspecific (e.g. I showed it a jar, a coffee cup, a glass, and even a plant pot and it classfied all of them simply as “cup”. While it is still fascinating that it did classify them at all I dont know why but I was expecting a little more specificity.
• The ml5 sketch that used MobileNet on the other hand was much more specific and could actually distinguish between these sorts of “cups” and while the guesses were not always perfect and it was a littl emore sporadic it was nice to see that it could at least tell me a little more about the objects it detected compared to the tensorFlow.js using the COCO dataset.
• what surprised me the most was that changing the scale of the objects (e.g. giving more background context to reference scale against or moving the object up close did not change the outcome of the classification too much.
• For sure both datasets lead to some pretty funny classifications, such as my plant pot with a cactus in it being labeled a “toilet seat” with the MobileNet dataset model and a whole acoustic guitar being labeled a “headband” with 67% certainty by the model using the COCO dataset.

TeachableMachine Making an interactive book with audio for specific pages

Week1.mp4

https://editor.p5js.org/FabriGu/sketches/BUVm1M78u ← Here is the link to the sketch though you need the book to be able to actually interact with it