Response to Excavating AI: The Politics of Images in Machine Learning Training Sets and Humans of AI
As someone who just started to understand the general architecture of AI models, the discussion on the political bias of datasets was illuminating. Despite the recurring hallucinations and errors, the current zeitgeist attributes an almost divine omnipotence to LLMs. I think this is perpetuated by three main forces.
The first is that for most people it’s been useful, and that is a pretty rare attribute for technology nowadays. The second is successful marketing that is creating hype worth billions of private and public sectors investments. The third is the inherent black box property of deep learning models. Wether you don’t know how GPT writes such accurate emails but don’t care as long as it does it for you, or a mathematician refining the algorithm without knowing which features it is using, much is unknown, and it just feels a bit like magic.
This essay successfully dispels much of the magic and even gets you thinking about how the rabbit ended up in that hat in the first place. This is important, because bias is most dangerous when it is disguised as objectivity. It is critical for the public to understand that AI models aren’t just wrong because there was an error, but because they are created in our image, and at least for the mean time, a fallible creator can’t create an omnipotent machine.
Emoji Mask by iss.dean -p5.js Web Editor
Train the teachable machine to recognize and label my facial expressions and display a corresponding emoji mask over a webcam feed in p5.js.
Step 1: Encode relevant sketch from https://editor.p5js.org/ml_4_cc/sketches/UeAjLnXQz
Step 2: Train teachable machine
Step 3: Write p5 sketch that connects to the ML model
Step 4: Write p5 function that traces the size and position of the face, and displays an emoji according to the facial expression
Step 1:
What does this line mean?
ml5.setBackend('webgl');
GPT explained that it sets the computational backend to webgl, which is basically the Javascript API for generating graphics with the computers GPU instead of CPU.