out of distribution
adjective, machine learning
- Of an input: unlike the data a model was trained on; outside its familiar range.
- Of a person: the same, said with admiration.
Engineers keep a phrase for the unfamiliar. It can show up in incident reports, said with a sigh, when the world stops resembling the examples. Little in the training data prepared the system for what just walked in.
A model trained on past examples can still meet something it does not know how to place. The phrase carries a small possibility: you were not implied by the past.
This page brought a small model. It has one job: to guess which button you will press, before you press it. It learns from every press. A rhythm, a repeated run, a habit you did not know you had. When it has too little to go on, it guesses.
The drawing keeps the evidence. A ring for its guess. A dot for your move. When the two part ways, a red line joins them. Your choices stay in this tab. Press “forget me” and its history is gone.
The score belongs to this small model and this sequence of choices. A person is a larger subject.
To be predicted is to be, in some small way, repeated. But the past is full of things someone did not see coming: a strange friendship, a sentence, a different way to spend a life. A model can meet something unfamiliar and return an error. We can meet the same thing and feel a beginning.
Same event, two readings. The machine files a failure report. We frame it.
May you always be slightly beyond the training data.