jonasloos/coupled-oscillators-mnist

This project is a small research experiment that generates images of handwritten digits, the kind found in the classic MNIST dataset, using a very different approach from typical AI image generators. Instead of the usual method of starting with noise and gradually removing it step by step, this system simulates a physical system of oscillators, things that swing back and forth like pendulums, that influence each other's timing. The final positions of these oscillators, once the simulation settles, are converted directly into an image. The core idea is that every single number the model learns is a parameter describing how these oscillators interact with each other, rather than describing image pixels directly. The step that turns the oscillator readings into a picture uses a fixed mathematical transform with no learned component at all, so the entire creative work of generating a recognizable digit is done by the physics simulation itself. To generate a specific digit, such as a 3 versus a 7, the model uses a separate small group of oscillators that carries information about which digit to draw, and the connections between that group and the main oscillators are what the model actually learns. The whole project lives in a single Python file and can be trained in minutes on an ordinary computer, using a training method based on a recent research technique that compares generated and real examples in a simplified feature space rather than using a more common adversarial training setup. According to the results reported in the README, the model can correctly guide toward a target digit about 86 percent of the time, though the generated digits look somewhat blurry or blob-like compared to more typical models, which the author attributes to the limited resolution of the fixed conversion step. The author notes this is a simplified, smaller-scale version of a similar approach originally applied to more complex image datasets.