Assignment 2b

We label images and categorize images all the time. And not only images, we label and categorize everything, everything we see, everything we touch, everything we listen to. Labeling is the core part of how we recognize thing. When we are trying to learn what is an apple, we are shown with an apple or an image of it and are taught that thing with shape like this is named apple. Ever since then when we see an object with such appearance, we categorize it as the apple. We make stereotypes when we only rely on labels in our head. We can't see the whole picture if we only depend on what we see. The labeling may be a false representation of the object's essence, like what Magritte claims " this is not a pipe". Trying to put something unfamiliar into the category we are familiar with is dangerous and bias. Similarly process happens to machine learning, same with the potential danger. Like how we learn the world, machine learning model are trained with training set, a whole set of images and their labels. AI then can analyze brand new images and put them into categories it encounters in the training process. All the recognition and decision AI made are fully built upon the data set, the label it is trained with, the label the human trainers create. The label is the trainer's own version of representation. The information a single label delivers is incomplete, or even deceptive. And such biases are greatly disguised under the "neutral", "unpolitical" facade of the machine, computer, AI. I really like a line in the excavating AI "The whole endeavor of collecting images, categorizing them, and labeling them is itself a form of politics, filled with questions about who gets to decide what images mean and what kinds of social and political work those representations perform." Who creates the dataset, who makes the label, who has the power to categorize, dictates the action of AI.

When we make a misjudgement due to our reliance on preset labeling and experience, we can apologize and start over. We can refresh the categories in our head or do better - think more before making judgement. But machine learning model escalates personal bias to a whole another level. A machine learning model before being pointed out its bias nature can be downloaded and used many many times. Perhaps it has already been used in a real project and its unrightful result has already caused harm. AI nowadays is employed in many aspects of our life and the fact that it is being so wildly used means that we sometimes don't even notice its existence. It with a bias dataset is like a bomb. We don't know when it will explode.

Link to my p5js sketch: https://editor.p5js.org/robotproject/sketches/Ft_K8KyNf

For this project I want to create a machine learning model that detect "yes" or "no" . When I first train my model, I used nodding and shaking my head to represent yes or no. But the machine can't accurately distinguish the difference between this two because they are moving motion and have similar frames where my head is at the center, facing forward.

Also I found out that I forgot to create a control group, the background image which makes the learning more difficult. In my second try I decided to use static hand gesture to represent yes or no. I first create the background control group first with my facing looking directly at the camera without any hand gesture. I then use the image of my thumb standing up representing yes and two hands cross representing no.

This time the model runs much more accurately. So I upload it to p5js model and test it out there. When I found out that it works pretty well, I changes several features in P5 such as hiding the camera and adding the line "Can I ask you a question". If the detected label is "yes", a thank you will pop up; else, the "can i ask you a question" will stay forever there.