This article has greatly inspired me by comparing Magritte's "The Treachery of Images" with AI datasets. The article points out that Magritte's painting "This is Not a Pipe" warns us that there is a huge gap between images and reality, and that meaning should be fluid and reinterpretable. However, datasets like ImageNet stand in opposition to Magritte. They arrogantly inherit the pseudo-scientific logic of 19th-century "physiognomy", superstitiously believing that human essence - whether it be personality or class - is "inscribed" in photos through biological features and can be precisely measured by machines through pixels. This makes me deeply think of the "phonetic physiognomy" in the reading material of my other ITP course, "Listening Machine", which is exactly the same as the logic of ImageNet. Both are trying to infer people's social attributes (class, personality, morality) through physical data (sound waves/pixels). They are simply the visual and auditory versions of the same kind of "modern witchcraft". If ImageNet is trying to judge a person's morality by "reading faces" (such as the labels "loser" and "alcoholic"), then "Listening Machine" is trying to distinguish people's social status by "listening to voices". They both commit a serious "category mistake", attempting to forcibly bind sociological abstract evaluations to physiological features. This technology is essentially a violent resurgence of "biological determinism". In this algorithmic era, the crisis we face is not only the leakage of privacy, but also the loss of the right to define. The current logic of AI is philosophically very naive and regressive. It does not understand "metaphor", does not understand "irony", and regards images as absolute evidence. This can be reflected upon: we need an AI that is more like an "artist" rather than one that is more like a "policeman". When AI starts to define who we are with these biased "old dictionaries", we must remain vigilant like Magritte: this is not only not a pipe, but even less an objective truth.
https://editor.p5js.org/yz10444/sketches/UuxOWKQSJ
I want to make a p5js game that can be controlled by gestures, because games that recognize through the camera often have a stronger sense of immersion.
In addition, considering that the model's recognition is not so stable and precise and there might be a delay, I set the size of the obstacles and the player (user controlled) to increase the tolerance. The overall experience is quite good.
Spreading out the five fingers means "up", and clenching them into a fist means "down".
model：https://teachablemachine.withgoogle.com/models/ah8ROjs1p/