Reading Excavating AI confirmed some fears and assumptions I had about assignment 1B. Learning about how ImageNet ripped a bunch of images off of Google Images to use in training without any sort of regard whatsoever as to the context of those images and how it ended up classifying ones involving people into twisted and outright offensive categorizations was something new, however. I was aware of the possibilities that AI could be trained to have bias, but I did not expect it to categorize things like this. Something I felt that I didn’t quite understand was how it exactly came to these wild conclusions and categorizations, and if it did actually take very old data from people back when photography was first a thing.

I think this passage did well to tell me to tread carefully with these image models, especially in what I would use them for. Like I wouldn’t exactly be comfortable trusting my safety in the hands of a machine which makes all these offensive and bizarre categorizations for people (the guy who used an image classifier on a drone to make it detect potential threats).

https://editor.p5js.org/Andr0543/sketches/zJDf2dRer

I decided to make a sort of “attention span” detector, where it detects how much you’re looking at the screen. If you look away for long enough, it plays a sound to get your attention.

https://drive.google.com/file/d/1OJRZYRkToztEMZ9B5ajCCmMQbdFlkwkk/view?usp=sharing

I trained the model using Teachable Machine with 3 classes: Focused, Drifting, and Absent, depending on how much you were looking away.

I used p5.sound and millis() to start a counter which would reset every time it detected “Focused” or “Drifting” instead of fully “Absent”. After 5 seconds, the program gets your attention with a “huh?” sound.