Machine Learning

Well, that's a correct statement. The only nondeterministic things that today's computers do, strictly speaking, are from one of two places:

• Hardware failures
• True random entropy hardware inputs (e.g. a hardware sensor recording noise in mouse or keyboard actions, hardware background noise, EM field strength, etc, or even better detecting things like random cosmic ray muon decays)

If we narrow our definition of nondeterministic to the PoV of a single process running on a multi-tenant machine, then there is more, coming from things like OS-controlled thread processing timings that the application has no visibility into.

But today's ML algorithm outputs are generally deterministic, outside hardware failure, and after you decide on random initialization seeds. They aren't drawing from random entropy during computation invisibly.

The model weights, input, system memory and storage state and random seeds completely determine the output, generally.

Well every computer has a chance due to physical interaction to make a random error in any algorithm.

So for sure there must be a random element in the output of the model as long as you calculate it on real hardware.

If you keep Dropout at Inference time, you don't get deterministic results, even if the input stays constant. People sometimes think you can use this to derive uncertainty (I don't).